Download the Full Version of textbook for Fast Typing at textbookfull.

com

Engineering Applications of Neural Networks 19th

International Conference EANN 2018 Bristol UK
September 3 5 2018 Proceedings Elias Pimenidis

https://textbookfull.com/product/engineering-applications-
of-neural-networks-19th-international-conference-
eann-2018-bristol-uk-september-3-5-2018-proceedings-elias-
pimenidis/

OR CLICK BUTTON

DOWNLOAD NOW

Download More textbook Instantly Today - Get Yours Now at textbookfull.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

Engineering Applications of Neural Networks 17th

International Conference EANN 2016 Aberdeen UK September 2
5 2016 Proceedings 1st Edition Chrisina Jayne
https://textbookfull.com/product/engineering-applications-of-neural-
networks-17th-international-conference-eann-2016-aberdeen-uk-
september-2-5-2016-proceedings-1st-edition-chrisina-jayne/
textboxfull.com

Computational Collective Intelligence 10th International

Conference ICCCI 2018 Bristol UK September 5 7 2018
Proceedings Part I Ngoc Thanh Nguyen
https://textbookfull.com/product/computational-collective-
intelligence-10th-international-conference-iccci-2018-bristol-uk-
september-5-7-2018-proceedings-part-i-ngoc-thanh-nguyen/
textboxfull.com

Computational Collective Intelligence 10th International

Conference ICCCI 2018 Bristol UK September 5 7 2018
Proceedings Part II Ngoc Thanh Nguyen
https://textbookfull.com/product/computational-collective-
intelligence-10th-international-conference-iccci-2018-bristol-uk-
september-5-7-2018-proceedings-part-ii-ngoc-thanh-nguyen/
textboxfull.com

Collaborative Networks of Cognitive Systems 19th IFIP WG 5

5 Working Conference on Virtual Enterprises PRO VE 2018
Cardiff UK September 17 19 2018 Proceedings Luis M.
Camarinha-Matos
https://textbookfull.com/product/collaborative-networks-of-cognitive-
systems-19th-ifip-wg-5-5-working-conference-on-virtual-enterprises-
pro-ve-2018-cardiff-uk-september-17-19-2018-proceedings-luis-m-
camarinha-matos/
textboxfull.com
Engineering Applications of Neural Networks: 20th
International Conference, EANN 2019, Xersonisos, Crete,
Greece, May 24-26, 2019, Proceedings John Macintyre
https://textbookfull.com/product/engineering-applications-of-neural-
networks-20th-international-conference-eann-2019-xersonisos-crete-
greece-may-24-26-2019-proceedings-john-macintyre/
textboxfull.com

Reversible Computation 10th International Conference RC

2018 Leicester UK September 12 14 2018 Proceedings Jarkko
Kari
https://textbookfull.com/product/reversible-computation-10th-
international-conference-rc-2018-leicester-uk-
september-12-14-2018-proceedings-jarkko-kari/
textboxfull.com

Lifelong Technology-Enhanced Learning: 13th European

Conference on Technology Enhanced Learning, EC-TEL 2018,
Leeds, UK, September 3-5, 2018, Proceedings Viktoria
Pammer-Schindler
https://textbookfull.com/product/lifelong-technology-enhanced-
learning-13th-european-conference-on-technology-enhanced-learning-ec-
tel-2018-leeds-uk-september-3-5-2018-proceedings-viktoria-pammer-
schindler/
textboxfull.com

Electronic Government and the Information Systems

Perspective: 7th International Conference, EGOVIS 2018,
Regensburg, Germany, September 3–5, 2018, Proceedings
Andrea K■
https://textbookfull.com/product/electronic-government-and-the-
information-systems-perspective-7th-international-conference-
egovis-2018-regensburg-germany-september-3-5-2018-proceedings-andrea-
ko/
textboxfull.com

Artificial Neural Networks and Machine Learning ICANN 2018

27th International Conference on Artificial Neural
Networks Rhodes Greece October 4 7 2018 Proceedings Part I
V■ra K■rková
https://textbookfull.com/product/artificial-neural-networks-and-
machine-learning-icann-2018-27th-international-conference-on-
artificial-neural-networks-rhodes-greece-october-4-7-2018-proceedings-
part-i-vera-kurkova/
textboxfull.com
Elias Pimenidis
Chrisina Jayne (Eds.)

Communications in Computer and Information Science 893

Engineering Applications
of Neural Networks
19th International Conference, EANN 2018
Bristol, UK, September 3–5, 2018
Proceedings

123
Communications
in Computer and Information Science 893
Commenced Publication in 2007
Founding and Former Series Editors:
Phoebe Chen, Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu,
Dominik Ślęzak, and Xiaokang Yang

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
Junsong Yuan
University at Buffalo, The State University of New York, Buffalo, USA
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at http://www.springer.com/series/7899
Elias Pimenidis Chrisina Jayne (Eds.)
•

Engineering Applications
of Neural Networks
19th International Conference, EANN 2018
Bristol, UK, September 3–5, 2018
Proceedings

123
Editors
Elias Pimenidis Chrisina Jayne
University of the West of England Oxford Brookes University
Bristol Oxford
UK UK

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-3-319-98203-8 ISBN 978-3-319-98204-5 (eBook)
https://doi.org/10.1007/978-3-319-98204-5

Library of Congress Control Number: 2018950649

© Springer Nature Switzerland AG 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

This CCIS Springer volume contains the proceedings of the 19th International Con-
ference on Engineering Applications of Neural Networks (EANN 2018), which was
held at the University of the West of England, Bristol, UK, during September 3–5,
2018. The conference was supported by the department of Computer Science and
Creative Technologies and the Faculty of Environment and Technology of the
University of the West of England.
With a history of 23 years since the inaugural event in Otaniemi, Finland, in 1995,
EANN has grown in reputation and has been attracting established as well as new
researchers in the field of neural networks and other related areas of artificial intelli-
gence. Since 2009, EANN has be supported technically by the International Neural
Network Society (INNS).
EANN 2018 attracted submissions from countries around the globe with the pre-
senting authors/delegates coming from 12 countries, namely, Belgium, China,
Germany, Italy, Japan, Norway, Oman, Pakistan, Portugal, Turkey, the USA, and the
UK. The volume includes 16 papers that were accepted for oral presentation as full
papers (at 12 pages each), representing a 41% acceptance rate on the number of sub-
missions. The Program Committee encouraged by the quality of submissions by young
researchers also accepted a further five submissions to be presented as short papers and
are included in this volume. All papers were subject to a rigorous peer-review process by
at least two independent academic referees, members of the international Program
Committee.
The accepted papers demonstrate a variety of novel neural network and other arti-
ficial intelligence approaches applied to challenging real-world problems. The papers
cover topics such as: activity recognition, deep learning, extreme learning machines,
fuzzy systems, machine learning applications, predictive models, recommender sys-
tems, recurrent neural networks, and spiking neural networks applications.
The following keynote speakers were invited and gave lectures on exciting neural
network application topics:
• Professor Plamen P. Angelov, Director of LIRA (Lancaster Intelligent, Robotic and
Autonomous systems) Research Centre, Lancaster University, UK
• Professor Anthony Piper, Deputy Director of the Bristol Robotics Laboratory,
UWE, Bristol, UK
• Professor Chrisina Jayne, Head of School of Engineering, Computing and Mathe-
matics, Oxford Brookes University, UK
On behalf of the conference Organizing Committee, we would like to thank all those
who contributed to the organization of this year’s program, and in particular the
Program Committee members.

September 2018 Elias Pimenidis

Chrisina Jayne
 Organization

General Chairs
Anthony Pipe University of the West of England, UK
Elias Pimenidis University of the West of England, UK

Organizing Chairs
Steve Battle University of the West of England, UK
Antisthenis Tsompanas University of the West of England, UK

Program Chairs
Chrisina Jayne Oxford Brookes University, UK
Phil Legg University of the West of England, UK

Advisory Chairs
Ilias Maglogiannis University of Piraeus, Greece
Lazaros Iliadis Democritus University of Thrace, Greece

Workshop Chairs
Mehmet Aydin University of the West of England, UK
Nikolaos Polatidis University of Brighton, UK

Publicity Chair
Plamen Angelov Lancaster University, UK

Honorary Committee
John MacIntyre University of Sunderland, UK
Nikola Kasabov Auckland University of Technology, New Zealand

Publications Chair
Antonios Papaleonidas Democritus University of Thrace, Greece
VIII Organization

Student Forum Chairs

Rong Yang University of the West of England, UK
Emmanuel Ogunshile University of the West of England, UK

Website Chair
Martyn Fitzgerald University of West of England, UK

Program Committee
Michel Aldanondo Toulouse University CGI, France
Athanasios Alexiou Novel Global Community Educational Foundation,
Australia
Ioannis Anagnostopoulos University of Thessaly, Greece
Plamen Angelov Lancaster University, UK
Mehmet Emin Aydin University of the West of England, UK
Costin Badica University of Craiova, Romania
Rashid Bakirov Bournemouth University, UK
Zbigniew Banaszak Warsaw University of Technology, Poland
Steve Battle University of the West of England, UK
Bartlomiej Beliczynski Warsaw University of Technology, Poland
Kostas Berberidis University of Patras, Greece
Nik Bessis Edge Hill University, UK
Monica Bianchini Università di Siena, Italy
Giacomo Boracchi Politecnico di Milano, Italy
Farah Bouakrif University of Jijel, Algeria
Anne Canuto Federal University of Rio Grande do Norte, Brazil
Diego Carrera Politecnico di Milano, Italy
George Caridakis National Technical University of Athens, Greece
Ioannis Chamodrakas National and Kapodistrian University of Athens,
Greece
Aristotelis Chatziioannou National Hellenic Research Foundation, Greece
Ben Daubney MBDA, UK
Jefferson Rodrigo De Souza FACOM/UFU, Brazil
Kostantinos Demertzis Democritus University of Thrace, Greece
Ioannis Dokas Democritus University of Thrace, Greece
Ruggero Donida Labati Politecnico di Milano, Italy
Mauro Gaggero National Research Council of Italy, Italy
Christos Georgiadis University of Macedonia, Greece
Giorgio Gnecco IMT - Institute for Advanced Studies, Lucca, Italy
Denise Gorse University College London, UK
Foteini Grivokostopoulou University of Patras, Greece
Xiaowei Gu Lancaster University, UK
Hakan Haberdar University of Houston, USA
Petr Hajek University of Pardubice, Czech Republic
 Organization IX

Ioannis Hatzilygeroudis University of Patras, Greece

Martin Holena Institute of Computer Science, Czech Republic
Amin Hosseinian-Far University of Northampton, UK
Chrisina Jayne Oxford Brookes University, UK
Jacek Kabzinski Lodz University of Technology, Poland
Antonios Kalampakas Democritus University of Thrace, Greece
Ryotaro Kamimura Tokai University, Japan
Dmitry Kangin University of Exeter, UK
Stelios Kapetanakis University of Brighton, UK
Ioannis Karydis Ionian University, Greece
Petros Kefalas The University of Sheffield, UK
Muhammad Khurram Khan King Saud University, Saudi Arabia
Kyriaki Kitikidou Democritus University of Thrace, Greece
Avrilia Kogketsof Democritus University of Thrace, Greece
Yiannis Kokkinos University of Macedonia, Greece
Mikko Kolehmainen University of Eastern Finland, Finland
Petia Koprinkova-Hristova Bulgarian Academy of Sciences, Bulgaria
Konstantinos Koutroumbas National Observatory of Athens, Greece
Paul Krause University of Surrey, UK
Ondrej Krejcar University of Hradec Kralove, Czech Republic
Efthyvoulos Kyriacou Frederick University, Cyprus
Phil Legg University of the West of England, UK
Florin Leon Gheorghe Asachi Technical University of Iasi,
Romania
Aristidis Likas University of Ioannina, Greece
Spiros Likothanassis University of Patras, Greece
Ilias Maglogiannis University of Piraeus, Greece
George Magoulas Birkbeck College, UK
Mario Natalino Malcangi Università degli Studi di Milano, Italy
Francesco Marcelloni University of Pisa, Italy
Konstantinos Margaritis University of Macedonia, Greece
Nikolaos Mitianoudis Democritus University of Thrace, Greece
Valeri Mladenov Technical University Sofia, Greece
Haralambos Mouratidis University of Brighton, UK
Phivos Mylonas National Technical University of Athens, Greece
Stavros Ntalampiras Politecnico di Milano, Italy
Mihaela Oprea University Petroleum-Gas of Ploiesti, Romania
Elpiniki Papageorgiou Technological Educational Institute of Lamia, Greece
Basil Papadopoulos Democritus University of Thrace, Greece
Antonios Papaleonidas Democritus University of Thrace, Greece
Michalis Pavlidis University of Brighton, UK
Daniel Pérez University of Oviedo, Portugal
Isidoros Perikos University of Patras, Greece
Miltos Petridis Middlesex University London, UK
Elias Pimenidis University of the West of England, UK
Anthony Pipe University of the West of England, UK
X Organization

Vassilis Plagianakos University of Thessaly, Greece

Nikolaos Polatidis University of Brighton, UK
Bernardete Ribeiro University of Coimbra, Portugal
Alexander Ryjov Moscow State University, Russia
Araceli Sanchis Universidad Carlos III de Madrid, Spain
Marcello Sanguineti Universita di Genova, Italy
Simone Scardapane Sapienza University of Rome, Italy
Olga Senyukova Moscow State University, Russia
Stephanos Spartalis Democritus University of Thrace, Greece
Ricardo Tanscheit Pontifical Catholic University of Rio de Janeiro, Brazil
Teck Hou Teng Singapore Management University, Singapore
Francesco Trovo Politecnico di Milano, Italy
Athanasios Tsadiras Aristotle University of Thessaloniki, Greece
Nicolas Tsapatsoulis Cyprus University of Technology, Cyprus
George Tsekouras University of the Aegean, Greece
Aristeidis Tsitiridis Rey Juan Carlos University, Spain
Michail - Antisthenis University of the West of England, UK
Tsompanas
Giannis Tzimas Technological Educational Institute of Western Greece,
Greece
Theodoros Tzouramanis University of the Aegean, Greece
Nikos Vassilas T.E.I. of Athens, Greece
Petra Vidnerová Institute of Computer Science, AS CR, Czech Republic
Panagiotis Vlamos Ionian University, Greece
George Vouros University of Piraeus, Greece
Xin-She Yang Middlesex University London, UK
Kostantinos Yalouris Agricultural University of Athens, Greece
Drago Žagar University of Osijek, Croatia
 Contents

Activity Recognition

A Framework for Semi-Supervised Adaptive Learning for Activity

Recognition in Healthcare Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Prankit Gupta and Praminda Caleb-Solly

Structured Inference Networks Using High-Dimensional Sensors

for Surveillance Purposes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Vincent Polfliet, Nicolas Knudde, Baptist Vandersmissen,
Ivo Couckuyt, and Tom Dhaene

Deep Learning

Deep Imitation Learning with Memory for Robocup Soccer Simulation . . . . . 31

Ahmed Hussein, Eyad Elyan, and Chrisina Jayne

Toward Video Tampering Exposure: Inferring Compression

Parameters from Pixels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Pamela Johnston, Eyad Elyan, and Chrisina Jayne

RR-FCN: Rotational Region-Based Fully Convolutional Networks

for Object Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Dingqian Zhang, Hui Zhang, Haichang Li, and Xiaohui Hu

Face Detection for Crowd Analysis Using Deep Convolutional

Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Bryan Kneis

Extreme Learning Machine and Machine Learning Applications

Smoothing Regularized Extreme Learning Machine . . . . . . . . . . . . . . . . . . . 83

Qin-Wei Fan, Xing-Shi He, and Xin-She Yang

Neuroevolution of Actively Controlled Virtual Characters - An Experiment

for an Eight-Legged Character . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Svein Inge Albrigtsen, Alexander Imenes, Morten Goodwin, Lei Jiao,
and Vimala Nunavath

Machine Learning with the Pong Game: A Case Study . . . . . . . . . . . . . . . . 106

Benedikt Nork, Geraldine Denise Lengert, Robert Uwe Litschel,
Nasim Ahmad, Gia Thuan Lam, and Doina Logofătu
XII Contents

Managing Congestion in Vehicular Networks Using Tabu Search . . . . . . . . . 118

Muhammad Ishaq, Mazhar Hussain Malik, and Mehmet Emin Aydin

Network Intrusion Detection on Apache Spark with Machine

Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Elif Merve Kurt and Yaşar Becerikli

Predictive Models, Fuzzy and Recommender Systems

A Triangle Multi-level Item-Based Collaborative Filtering Method

that Improves Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Gharbi Alshammari, Stelios Kapetanakis, Nikolaos Polatidis,
and Miltos Petridis

Myo-To-Speech - Evolving Fuzzy-Neural Network Prediction

of Speech Utterances from Myoelectric Signals. . . . . . . . . . . . . . . . . . . . . . 158
Mario Malcangi, Giovanni Felisati, Alberto Saibene, Enrico Alfonsi,
Mauro Fresia, Roberto Maffioletti, and Hao Quan

Model Prediction of Defects in Sheet Metal Forming Processes . . . . . . . . . . 169

Mario Dib, Bernardete Ribeiro, and Pedro Prates

Selecting Display Products for Furniture Stores Using Fuzzy Multi-criteria

Decision Making Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Özer Uygun, İlker Güven, Fuat Şimşir, and Mehmet Emin Aydin

Reproduction of Experiments in Recommender Systems Evaluation

Based on Explanations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Nikolaos Polatidis and Elias Pimenidis

Recurrent Neural Networks and Spiking Neural Networks

Recurrent Auto-Encoder Model for Large-Scale Industrial

Sensor Signal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Timothy Wong and Zhiyuan Luo

Deep Neural Networks for Prediction of Exacerbations of Patients

with Chronic Obstructive Pulmonary Disease . . . . . . . . . . . . . . . . . . . . . . . 217
Vimala Nunavath, Morten Goodwin, Jahn Thomas Fidje,
and Carl Erik Moe

Probabilistic Word Association for Dialogue Act Classification

with Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Nathan Duran and Steve Battle
 Contents XIII

Acceleration of Convolutional Networks Using Nanoscale

Memristive Devices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Shruti R. Kulkarni, Anakha V. Babu, and Bipin Rajendran

Comparison of Asymmetric and Symmetric Neural Networks

with Gabor Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Naohiro Ishii, Toshinori Deguchi, Masashi Kawaguchi,
and Hiroshi Sasaki

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265

Activity Recognition
 A Framework for Semi-Supervised Adaptive
Learning for Activity Recognition in Healthcare
Applications

Prankit Gupta ✉ and Praminda Caleb-Solly

( )

Bristol Robotics Laboratory, University of the West of England, Bristol BS16 1QY, UK
[email protected], [email protected]

Abstract. With the growing popularity of the Internet of Things and connected
home products, potential healthcare applications in a smart-home context for
assisted living are becoming increasingly apparent. However, challenges in
performing real-time human activity recognition (HAR) from unlabelled data and
adapting to changing user health remain a major barrier to the practicality of such
applications. This paper aims to address these issues by proposing a semi-super‐
vised adaptive HAR system which combines offline and online recognition tech‐
niques to provide intelligent real-time support for frequently repeated user activ‐
ities. The viability of this approach is evaluated by pilot testing it on data from
the Aruba CASAS dataset, and additional pilot data collected in the Bristol
Robotics Lab’s Assisted Living Studio. The results show that 71% of activity
instances were discovered, with an F1-score of 0.93 for the repeating
“Meal_Prep” activities. Furthermore, real-time recognition on the collected pilot
data occurred near the beginning of the activity 64% of the time and at the halfway
point in the activity 96% of the time.

Keywords: Activity recognition · Smart-home · Healthcare

1 Introduction

The term smart-home (SH) refers to the concept of integrating everyday items/home
appliances with various sensors, actuators, and relays, and their inter-networking to
achieve various forms of automation [1]. An increasing population of older adults with
ageing-related impairments and long-term conditions is putting a greater amount of
pressure on healthcare providers, who are struggling to balance high-quality care with
reduced budgets. The utilization of IoT technology within a SM context to provide
Ambient Assisted Living (AAL) is regarded as a key solution to achieve this balance [2].
An important area of research in SHs for AAL is human activity recognition (HAR).
Offline and online HAR are two different approaches for detecting activities, with online
methods working in real-time as the activity is occurring [3]. Most conventional HAR
approaches that have shown promising results in the past have used supervised learning,
requiring large amounts of expert-labelled user activity training data which is both
difficult to obtain and not practical for real-world deployment [4]. These systems also

© Springer Nature Switzerland AG 2018

E. Pimenidis and C. Jayne (Eds.): EANN 2018, CCIS 893, pp. 3–15, 2018.
https://doi.org/10.1007/978-3-319-98204-5_1
4 P. Gupta and P. Caleb-Solly

largely remain unable to adapt to changing user health and behaviour over time, as they
require supervised re-training with new activity data to account for this.
This paper proposes an alternative approach to conventional HAR, where unlike
traditional systems the focus is not on training a classifier to identify every instance of
every activity, but rather on discovering frequently repeating activity patterns. Accord‐
ingly, this paper presents a framework which utilizes a pre-defined short-term system
memory for performing offline HAR on unlabelled data and updating the user activity
model using Bayesian Networks for real-time detection. This combination of offline and
online HAR works simultaneously to constantly learn activity patterns and evolve the
user activity model on a short-term basis to account for any new or changing activities.
This evolving user activity model could be used to set-up automated prompts for users
with cognitive decline as well as track and continually assess user health to complement
carer support. An implementation of this approach is also pilot tested with the Aruba
CASAS dataset [5] and pilot data collected in the Assisted Living Studio (ALSt) which
is set up as a realistic home environment in the Bristol Robotics Laboratory.
The rest of the paper is divided as follows – Sect. 2 provides background information
and outlines related work; Sect. 3 formulates the problem; Sect. 4 presents the overall
framework for the HAR system, including descriptions of the offline and online HAR
systems and machine learning methods used; Sect. 5 describes the data sets used for
pilot testing the approach; Sect. 6 outlines the procedure followed in developing and
testing the system; Sect. 7 presents the results and related discussion; and finally
Sect. 8 summarises the conclusions and discusses future work.

2 Background and Related Work

This section reviews existing research to establish the current state of the art in SH sensor
systems, as well as determine the strengths and limitations of the machine learning
algorithms and adaptive environments used.

2.1 Smart-Home Sensor Systems

SHs can consist of a variety of different sensors ranging from ambient sensors, body-sensor
networks, to video-based solutions [6]. Ambient wireless sensor systems consist of sensors
embedded in the environment of the user such as Passive Infrared (PIR) motion, magnetic
contact, temperature, and LUX sensors [7]. Generally, a large number of ambient sensors
are required to be present in the room in order to track an activity [8], however they are
viewed as less intrusive and more acceptable than other types of sensing techniques which
include video monitoring [9]. Body-sensor systems include devices that are physically
worn or carried by the user such as fitness monitors and smartphones [10]. These can
provide physiological data such as heart-rate and body temperature, along with activity
data through embedded accelerometer, gyroscope, etc. The downside of body sensors is
that they are often viewed as intrusive and older adults may forget to wear them every day
[6]. Video-based methods generally provide the most contextual information for HAR, but
are viewed as less acceptable due to security and privacy issues [11].
 A Framework for Semi-Supervised Adaptive Learning 5

2.2 Machine Learning Algorithms for Activity Recognition Systems

Offline Learning HAR systems
In activity recognition, semi-supervised and unsupervised algorithms are generally
deployed for discovering activities from unlabelled data in an offline system. Kwon et al.
[12] present a comparison of unsupervised clustering techniques which can be used to
discover activity clusters in unlabelled sensor data. Their research focused on recog‐
nising activities from smartphone data such as walking, sitting, standing and running.
The unsupervised machine learning techniques included Gaussian mixture models,
Hierarchical Agglomerative Clustering (HAC), DBSCAN, and K-means clustering.
HAC and DBSCAN achieved the highest accuracy and the authors noted HAC for its
flexibility and ability to discover hierarchies in the data. HAC works by initially
assigning each data point as a cluster, then merging the clusters together at each step. A
cut-off can be set as to when the merging should stop, and the various steps of the
clustering process can be viewed in the form of a dendrogram.

Real-time Learning HAR and Adaptive Environments

Yala et al. [13] present two methods that utilize an incremental support vector machine
(SVM) along with clustering techniques to perform online HAR from streaming sensor
data. They utilize a sliding sensor window that encodes the preceding sensor events in
order to classify them. Other researchers have also experimented with similar sliding
window techniques, ranging from static length sensor windows to dynamic time
windows [14].
Drawbacks of traditional supervised HAR techniques are well summarised in [15],
in which Ntalampiras and Roveri propose an interesting framework that utilizes Hidden
Markov Models (HMM) for performing HAR and incremental dictionary learning to
account for new user activities. The only assumption made in this framework is that the
activities last for a specific amount of time. The researchers evaluated their technique
on gyroscope and accelerometer data from activities such as walking, sitting and
standing with promising results.
Most HAR approaches reviewed have utilized supervised learning or have been
focused on simple physical activities. This paper instead proposes a semi-supervised
framework for recognising and modelling more complex activities of daily living such
as cooking, whilst also adapting to changing user behaviour.

3 Problem Formulation

There are two problems the research presented in this paper aims to solve – performing
HAR from unlabelled data and tracking user activity in real-time. Traditional HAR
approaches train a classifier on datasets which include a labelled dictionary of activities:
 = {A1,…AN}, where Ai denotes the i-th activity and N is the total number of activities.
The classifier is then utilized to classify new occurrences of activities as one that is
present in . This method has shown promising results, but does not account for any
new activities that the user may start performing which were not present in the original
dictionary or changes over time in the way the user performs those activities.
6 P. Gupta and P. Caleb-Solly

The second problem involves tracking the user activity which is crucial to providing
real-time support. To formulate this problem, we first define the term user activity (or
task) as a combination of multiple sets of user actions (or subtasks): A = {S1,…SN}, where
Si denotes the i-th subtask and N is the total number of subtasks. The sequence in which
these subtasks are performed may vary, as there may be multiple routes to the end goal
of the task [16]. The challenge involved in tracking user activity is modelling the rela‐
tionships between subtasks present in the activity, in order to predict the next sequence
of subtasks.
Therefore, the goal of this research is to present a framework which can - automat‐
ically generate a dictionary of frequently repeated activities (defined in the next section)
by analysing unlabelled sensor data; modify this dictionary according to changes in user
behaviour; and model the relationships between the subtasks for each activity to allow
for real-time tracking and support.

4 Adaptive Activity Recognition Framework

This section describes the overall framework of the proposed HAR system which aims
to identify and model frequently repeated activities performed by the user in order to
set up automated assistance/prompts and continually track changes in the performance
of these activities for health assessment purposes. The system utilizes ambient and
passive sensors embedded throughout the house rather than wearables or video data due
to their greater acceptability as established in the literature review.
We first introduce a short-term system memory for storing and analysing sensor data.
The system memory is specified as ‘n’ days and is the number of preceding days’s data
the system would use from which to identify activities. If the present day is p, then a
repeating activity will be identified as such if it is repeated at least once since day p-n.
The value of n can be set to a larger integer for tracking longer-term activity model
changes, but a smaller value of n favours a more adaptive system. A real-world imple‐
mentation would require multiple non-overlapping system memories of different sizes
to keep track of weekday and weekend behaviour separately (as these might be quite
different). This paper only evaluates HAR of weekday behaviour.
The HAR system is divided into two parts, offline HAR and online HAR. Each of
these parts are comprised of three steps (Fig. 1). The purpose of the offline HAR system
is to discover activity clusters present in unlabelled user data at the end of each day and
compare it to the activity clusters found in the previous n days to identify any repeating
activities. This is then used to build a model of each identified repeating activity which
can be used for performing online or real-time HAR for automated assistance/prompts
and updating the activity model based on user actions.
These steps are repeated every day for the database of activity models to be created
and updated. Any data older than the specified system memory is deleted, which results
in old activities/habits being removed from the database. Any newly discovered
repeating activities are added, while existing activity models are updated based on user
behaviour (Fig. 2).
 A Framework for Semi-Supervised Adaptive Learning 7

Fig. 1. Adaptive human activity recognition framework

Fig. 2. Identifying and modelling a new repeating user activity introduced on day ‘p-1’

The following section describes the datasets used to develop and evaluate the system
while Sect. 6 describes each step in more detail along with an explanation of the feature
8 P. Gupta and P. Caleb-Solly

selection and the artificial intelligence implemented for each of the steps in this frame‐
work.

5 Aruba Dataset and Pilot Data Collection in the Assisted Living

Studio

The Aruba CASAS study deployed multiple motion sensors in each room of an older
female participant’s house which yielded labelled activity data for a total of 7 months
[5, 17]. Four weeks of kitchen data (two weeks from two different months) were selected
at random from this dataset, split into two weeks each for training/developing and testing
the offline HAR presented in this paper. Kitchen data was selected so that it could be
augmented by additional complementary data collected in the Assisted Living Studio
(ALSt) kitchen in the Bristol Robotics Laboratory. Weekend data was removed as only
the weekday data was used in this study as explained in 3.1. While the presence of real-
world noise in the Aruba dataset provided the ecological validity for testing the offline
HAR component of this study, the sensor configuration lacked inclusion of other sensors
such as contact sensors on drawers/cupboards, rendering the data insufficient to test the
online HAR system presented in this paper. As such, a sample dataset was collected in
the ALSt in the Bristol Robotics Laboratory (BRL) from participants recruited to
perform typical kitchen-based activities, which served to augment the Aruba dataset. A
Z-wave sensor network using openHAB was deployed in the ALSt which included
Fibaro FGMS-001 motion sensors, Everspring SM810 magnetic contact sensors on
cupboards, and TKB TZ69E wall plugs (to act as a sensor for the kettle). A total of six
activity sessions were recorded from four participants on different days. Due to the nature
of sensors used in the ALSt data collection, the sensor values only consisted of ON or
OFF (0 or 1) values similar to the motion sensors deployed in the Aruba data collection.
The number of ambient sensors deployed in the ALSt were also similar to the number
of motion sensors present in the various rooms in the Aruba study to ensure compatibility
between the two datasets.
The non-scripted kitchen activity selected for this study was a combination of two
interleaved sub-activities - making a hot drink (such as tea, coffee or hot chocolate) and
preparing a sandwich. The participants were provided with various items and ingredients
appropriate for the selected activity and were asked to place these in the ALSt kitchen
in a manner they were most comfortable with. They had access to two cupboards, one
drawer, a counter-top, fridge, along with a kettle for boiling water and relevant utensils.
In the selected Aruba dataset there was a 30 to 60 min gap everyday between 16:00
and 17:00 after removing unlabelled activities, which was replaced by an ALSt partic‐
ipant’s activity session data (Fig. 3). This was done to introduce real-world noise and
data from the Aruba dataset, with the combined Aruba-ALSt data used for testing the
offline HAR system, and the ALSt activity used for testing the online HAR system.
 A Framework for Semi-Supervised Adaptive Learning 9

Fig. 3. ALSt activity inserted into Aruba CASAS dataset

6 Pilot Evaluation Study

This section describes an implementation of the adaptive HAR framework along with
the data processing and machine learning techniques used. The weekday period was set
to 5 days for pilot testing with the combined Aruba-ALSt data. Matlab was used for all
data processing.

Step I: Data Pre-Processing

The sensor data obtained from the user for the specified system memory was separated
by days and grouped by rooms –

Separation of Days. This was performed so that the sensor data for each day can be
processed separately, and then compared to that of the previous day to identify repeating
activities.

Grouping Sensors by Rooms. An assumption this paper made was that the activities are
restricted by room, which would mean that if a single activity took place over three
rooms (such as cleaning), it would be split into three activities (such as cleaning room
1, cleaning room 2, and cleaning room 3). The sensor data was separated and stored
accordingly.

Step II: Discovering User Activities - Hierarchical Agglomerative Clustering

The purpose of this step was to discover user activities present in the data at the end of
each day. This paper implements HAC to achieve this as used by [12], as it has been
proved to be useful for discovering hierarchies of subtasks present within a task. The
features used for clustering were the normalised sensor IDs and the timestamp at which
the activity occurred. The maximum variance in the data was along the time axis, so
HAC would cluster together data points that occurred within short time intervals. The
distance calculated between data points was Euclidean. The linkage method was set to
‘single’, so that the shortest distance was used when clustering. A general rule for the
cut off for HAC is to divide the square root of the number of data points by two [18].
After the analysis of the generated dendrograms from the training data, the cut-off
formula was changed to
√
no. of sensor events
Cutoff = (1)
1.5
10 P. Gupta and P. Caleb-Solly

Therefore, the cut-off was set dynamically for each day depending on the total
number of sensor events present.

Step III: Discovering Repeating Activities through Cluster Comparison

Once distinct activity clusters have been discovered, each activity cluster was compared
to clusters from the previous five days to identify repeating activities. There were two
features used for this –
1. The time of the activity cluster. Calculated as follows –

time of last sensor event − time of first sensor event

Activity time = (2)
2

This feature was used to compare the time at which the activity occurred on different
days. This means that an activity had to be repeated within the same time frame each
day for it to be considered a repeating activity (set as one hour in this study after
training data analysis). This is a shortcoming of this current implementation of the
Adaptive Activity Recognition Framework which could be overcome by a future
implementation of dynamic time warping for cluster comparison [19].
2. The total number of sensor events present in the activity cluster. This feature was
used to represent the sizes of the activity clusters for comparison. In this paper clus‐
ters greater than 0.75 and smaller than 1.67 of the original were considered similar
(thresholds set after analysis of training data).
Sensor data belonging to activity clusters identified as similar/repeating were stored
using a unique identifier to refer to the activity. The data from each identifier was then
used to build the activity model for that identifier in the next step.

Step IV: Building User Activity Model using Bayesian Networks

Sensor data from previous days’ of user activity were used to generate a Bayesian net
of the activity (Fig. 4). This involves identifying and collapsing repeating sequences of
sensor events from each previous day. Figure 4 encodes all possible sequences of events
the user has performed for the activity in the last five days. Days 3, 4 and 6 are empty
as the activity from those days can be reproduced using the sequence of events from
days 1, 2 and 5. The probability of the user performing the next sequence of events
depends on how many times the user has repeated the same sequence in the past.
 A Framework for Semi-Supervised Adaptive Learning 11

Fig. 4. Bayesian network for assisted living studio activity for participant B

Step V: Real-Time HAR Using Sliding Sensor Window

Once the Bayesian model is generated, a sliding window comprising of three sensor
events is used for performing real-time HAR, similar to [13] as shown in Fig. 5 (Note:
this is different from the system memory described in Sect. 3.1). This is achieved by
first locating the unique activity models from previous days present in the database
(generated in step IV) which occurred within an hour of present time. Then comparing
the most recent three sensor events (sliding window size = 3) to the first three sensor
events present in the activity models to identify the activity being performed.

Fig. 5. Sliding window for sensor events – numbered lines are sensor events, lettered boxes are
sliding windows of 3 sensor events.

Step VI: Updating User Activity Model through User Feedback

Following real-time recognition, the SH algorithm needs to track the user in to order
detect any change in behaviour and automate prompts accordingly. For this, a Matlab
script was created to use the Bayesian nets created in the previous step to predict the
next most probable move by the user. If the user misses a step, the script keeps track of
the user by jumping ahead to reach the step the user is at. However, the script also keeps
track of any missed steps in order to provide a prompt to the user to remind him of the
missed step. According to the user’s response of either correcting for the missed step by
doing specific missing action, responding to the prompt with text input, or ignoring the
prompt, the Bayesian net for the activity is updated accordingly.
12 P. Gupta and P. Caleb-Solly

7 Results and Discussion

HAC (Step II) was evaluated on 12 days of Aruba weekday data for kitchen activities
between the dates 3rd January 2011 to 18th February 2011 with weekend data removed,
as this study only evaluated the weekday system memory. Recognised activities were
separated into activities identified precisely and activities identified with noise. An
activity was classified as identified precisely if the number of sensor events in the
identified activity cluster were same as the number of events listed in the dataset. Some
activity clusters contained additional sensor events not part of the label which were
classified as “identified with noise” (Fig. 6). The accuracy was calculated as follows

Number of activities identified precisely

Accuracy = × 100 (3)
Total number of activities labelled in dataset

Fig. 6. “Identified with Noise” meal preparation activity, 18 Feb 2011

The overall accuracy of the HAC was 71% for the Aruba dataset in separating sensor
events into separate activity clusters. The system identified six separate “Meal Prep”
clusters as there were a maximum of six separate instances of the “Meal Prep” activity
present in the same day, which could signify different variations of the activity such as
preparing different types of meals, making a hot drink, etc.
Cluster comparison (Step III) was then performed to identify the same activities
repeating on different days. The system compared the six “Meal Prep” clusters from step
II to identify repeating instances of the activity present on other days with an F1-score
of 0.93, while the score for the identifying repeating “Wash Dishes” activity was 0.35.
For comparison, authors in [13] used the Aruba dataset to achieve an F1-score between
0.6 and 0.7 for “Meal Prep” activities and 0 for “Wash Dishes”. They concluded that
the classifier failed to identify the “Wash Dishes” activity due to imbalance in the dataset
because of the presence of unlabelled sensor events that dominates it. This also seems
to be the case here as the “Wash Dishes” activity was performed less consistently than
“Meal Prep” in the selected data and the system often classified unlabelled data as “Wash
Dishes” during cluster comparison, resulting in the low F1-score. It must also be noted
that authors in [13] used six weeks of Aruba dataset including training and testing, while
this paper only used 12 days for both, making the former’s approach more robustly
tested. The system also identified instances of unlabelled data as an activity which were
removed from the results as they cannot be verified due to missing labels (Fig. 7).
 A Framework for Semi-Supervised Adaptive Learning 13

Fig. 7. An unlabelled repeating activity cluster identified and isolated

The F1-score for the cluster comparison of the ALSt activity from the combined
Aruba-ALSt dataset was 0.96. Real-time activity recognition (Step V) for the ALSt
participant activity from day 2 onwards occurred at the beginning of the activity (3rd
sensor event) 64% of the time, and near the midway point (5th to 7th sensor event) 96%
of the time. On day 6, real-time activity recognition occurred at the beginning of the
activity 100% of the time for all participants. This is because the Bayesian activity model
improves each day utilizing the previous 5 days of data. It must also be noted that the
earliest real-time detection can only happen at the 3rd sensor event due to the size of the
sliding sensor window being three sensor events.
The last experiment was to evaluate how well the system can track the user to detect
missed steps using the Bayesian network and provide prompts (Step VI), which was
achieved by simulating the user forgetting to open/close the cabinet and the fridge during
the activity on day 5 and day 6. The system detected these missed steps in 75% of the
instances on day 5 and 87.5% of the instances on day 6.

8 Conclusions and Future Work

A significant problem with current HAR techniques is the requirement of large amounts
of expertly labelled training data for each specific end-user. This paper presented an
alternative approach to HAR which utilises a short-term system memory with specified
decay length to recognise repeating activities from unlabelled data and track them in
real-time. The user activity models generated by this system can be utilized to offer
additional support, for example, if the user forgets or misses a step in a known activity
pattern such as forgetting to turn off the hob while preparing a meal, the system can then
remind them through automated prompts. Additionally, this approach can be used to
track changes in user activities which could be an indication of deteriorating health
conditions and alert the carer of significant changes based on a personalised threshold.
An implementation of the proposed HAR framework was also pilot tested using the
Aruba dataset augmented with sample data collected in the ALSt in the Bristol Robotics
Laboratory with promising results. Future work will include testing this framework on
14 P. Gupta and P. Caleb-Solly

different datasets and attempting to improve the cluster comparison performance by

implementing techniques such as dynamic time warping [19].

References

1. Demiris, G., Hensel, B.K.: Technologies for an aging society: a systematic review of “smart
home” applications. Yearb Med. Inf. 3, 33–40 (2008)
2. Kvedar, J., Coye, M.J., Everett, W.: Connected health: a review of technologies and strategies
to improve patient care with telemedicine and telehealth. Health Aff. (Millwood) 33, 194–
199 (2014)
3. Shoaib, M., Bosch, S., Incel, O., Scholten, H., Havinga, P.: A survey of online activity
recognition using mobile phones. Sensors 15, 2059–2085 (2015)
4. Ranasinghe, S., Al MacHot, F., Mayr, H.C.: A review on applications of activity recognition
systems with regard to performance and evaluation. Int. J. Distrib. Sens. Netw. 12,
1550147716665520 (2016)
5. CASAS: Smart Home Projects, Washington State University (USA). http://casas.wsu.edu/
6. Pal, S., Feng, T., Abhayaratne, C.: Real-time recognition of activity levels for ambient assisted
living. In: 5th IEEE International Conference Consumer Electronics - Berlin ICCE-Berlin
2015, pp. 485–488 (2016)
7. Roggen, D., Förster, K., Calatroni, A., Tröster, G.: The adARC pattern analysis architecture
for adaptive human activity recognition systems. J. Ambient Intell. Humaniz. Comput. 4,
169–186 (2013)
8. Cardinaux, F., Bhowmik, D., Abhayaratne, C., Hawley, M.S.: Video based technology for
ambient assisted living: A review of the literature. J. Ambient Intell. Smart Environ. 3, 253–
269 (2011)
9. Basu, D., Moretti, G., Sen Gupta, G., Marsland, S.: Wireless sensor network based smart
home: Sensor selection, deployment and monitoring. In: 2013 IEEE Sensors Applications
Symposium, SAS 2013 – Proceedings, pp. 49–54 (2013)
10. Rashidi, P., Mihailidis, A.: A survey on ambient-assisted living tools for older adults. IEEE
J. Biomed. Heal. Inform. 17, 579–590 (2013)
11. Boise, L., Wild, K., Mattek, N., Ruhl, M., Dodge, H.H., Kaye, J.: Willingness of older adults
to share data and privacy concerns after exposure to unobtrusive home monitoring.
Gerontechnology 11, 428–435 (2013)
12. Kwon, Y., Kang, K., Bae, C.: Unsupervised learning for human activity recognition using
smartphone sensors. Expert Syst. Appl. 41, 6067–6074 (2014)
13. Yala, N., Fergani, B., Fleury, A.: Feature extraction and incremental learning to improve
activity recognition on streaming data. In: 2015 IEEE International Conference on Evolving
and Adaptive Intelligent Systems EAIS 2015, pp. 1–8 (2015)
14. Krishnan, N.C., Cook, D.J.: Activity recognition on streaming sensor data. Pervasive Mob.
Comput. 10, 138–154 (2014)
15. Ntalampiras, S., Roveri, M.: An incremental learning mechanism for human activity
recognition. In: 2016 IEEE Symposium Series on Computational Intelligence, pp. 1–6 (2016)
16. Storf, H., Kleinberger, T., Becker, M., Schmitt, M., Bomarius, F., Prueckner, S.: An event-
driven approach to activity recognition in ambient assisted living. In: Tscheligi, M., et al.
(eds.) AmI 2009. LNCS, vol. 5859, pp. 123–132. Springer, Heidelberg (2009). https://doi.org/
10.1007/978-3-642-05408-2_16
17. Cook, D.J.: Learning setting- generalized activity models for smart spaces. IEEE Intell. Syst.
27, 32–38 (2012)
 A Framework for Semi-Supervised Adaptive Learning 15

18. Duda, R.O., Hart, P.E., Stork, D.G.: Pattern Classification, 2nd edn. John Wiley & Sons,
Hoboken (2001)
19. Petitjean, F., Ketterlin, A., Gançarski, P.: A global averaging method for dynamic time
warping, with applications to clustering. Pattern Recogn. 44, 678–693 (2011)
 Structured Inference Networks
Using High-Dimensional Sensors
for Surveillance Purposes

Vincent Polfliet1 , Nicolas Knudde1(B) , Baptist Vandersmissen2 ,

Ivo Couckuyt1 , and Tom Dhaene1
1
Department of Information Technology, Ghent University - imec, Ghent, Belgium
[email protected]
2
Department of Electronics and Information Systems,
Ghent University - imec, Ghent, Belgium

Abstract. Video cameras are arguably the world’s most used sensors
for surveillance systems. They give a highly detailed representation of
a situation that is easily interpreted by both humans and computers.
However, these representations can lose part of their representational
value when being recorded in less than ideal circumstances. Bad weather
conditions, low-light illumination or concealing objects can make the
representation more opaque. A radar sensor is a potential solution for
these situations, since it is unaffected by the light intensity and can
sense through most concealing objects. In this paper, we investigate the
performance of a structured inference network on data of a low-power
radar device. A structured inference network applies automated feature
extraction by creating a latent space out of which the observations can be
reconstructed. A classification model can then be trained on this latent
space. This methodology allows us to perform experiments for both per-
son identification and action recognition, resulting in competitive error
rates ranging from 0% to 6.5% for actions recognition and 10% to 12%
for person identification. Furthermore, the possibility of a radar sensor
being used as a complement to a camera sensor is investigated.

Keywords: Structured inference network · Person identification

Action recognition · Indoor sensing · Micro-doppler · Sensor fusion

1 Introduction
In recent years, interest in autonomous surveillance systems grew considerably.
While these systems improved significantly, the primary sensors remained the
same. The dominance of video cameras in autonomous surveillance systems
can be explained by their fundamental strengths. They give a detailed high
dimensional representation of their environment, which is easily interpretable
by humans as well as computers. Moreover, reduction in price and higher res-
olutions kept driving their success. While the fundamental advantages of video
c Springer Nature Switzerland AG 2018
E. Pimenidis and C. Jayne (Eds.): EANN 2018, CCIS 893, pp. 16–27, 2018.
https://doi.org/10.1007/978-3-319-98204-5_2
 Structured Inference Networks for Surveillance Purposes 17

cameras were a big catalyst for the early development of surveillance systems,
their deficiencies are now holding them back. For some of these deficiencies, such
as recordings in bad weather conditions or low light environments, workarounds
can be found. Others such as concealing clothing are harder to deal with. In con-
trast, a radar sensor is unaffected by concealing clothing, bad weather conditions,
low-light environments and can be placed out of sight, behind a wall.
A radar is an active sensor that transmits an electromagnetic signal, which
is reflected by objects in its line of sight. Information about these objects is
then extracted out of these signals taking advantage of, e.g. the Doppler effect.
Moreover, individual moving parts of a person or object will each reflect their
own Doppler signal which are then summarized into a micro-Doppler (MD)
signature [3].
These signatures contain information about the movement of the target, pro-
viding a promising feature to differentiate between for example cars, bikers,
pedestrians or dogs. Another use for these MD signatures is to recognise differ-
ent actions, ranging from walking to sitting or boxing [10,11]. However, perhaps
the most challenging application is to differentiate individuals based on the way
they move, the so called gait-based identification. While there is a noticeable dif-
ference between how a dog and a human walks or how a person runs or sits, the
difference in the MD signature between two persons walking is more subtle. This
subject has been extensively researched, however, previous papers used a high-
power radar sensor with relatively simple scenarios. In this paper the data sets
are recorded using a low-power frequency modulated continuous wave (FMCW)
radar. This radar is a low-cost, power efficient and compact sensor suited for
indoor usage. However, the combination of a human’s low radar cross-section
and a low-power device poses a significant challenge for this study [4].
Two data sets are used for our experiments. The first uses the IDentifica-
tion with Radar (IDRad) benchmark, which is an extensive data set where the
main objective is to identify individuals moving randomly in a room [19]. An
additional data set is recorded where the main objective is to recognise differ-
ent actions. Previous studies applied either deep convolutional neural networks
(DCNN) [11,19] or clustering methods [9,20] to MD signatures. Both approaches
were successful by exploiting certain properties of the data. The DCNN tries to
take advantage of the spatial properties, along the time and velocity axes, of an
MD signature. Conversely, the clustering methods are applied on feature vectors
of the original noisy data. Hence, a structured inference network (SIN) [14] can
potentially exploit both these properties due to its inherent Markovian proper-
ties. This model creates a lower dimensional latent space into which each time
step is projected without losing their sequential dependencies. The lower dimen-
sional states also implies that the model performs autonomous feature selection
on the data. The resulting lower dimensional latent states are then used in a clas-
sification model. These properties make the SIN well-suited for high dimensional
sequential data, such as radar data.
Exploring the Variety of Random
Documents with Different Content
The Project Gutenberg eBook of Lord Lister
No. 0306: Een avontuur van Koning Alfonso
 This ebook is for the use of anyone anywhere in the United
States and most other parts of the world at no cost and with
almost no restrictions whatsoever. You may copy it, give it away
or re-use it under the terms of the Project Gutenberg License
included with this ebook or online at www.gutenberg.org. If you
are not located in the United States, you will have to check the
laws of the country where you are located before using this
eBook.

Title: Lord Lister No. 0306: Een avontuur van Koning Alfonso

Author: Theo von Blankensee

Felix Hageman
Kurt Matull

Release date: December 23, 2023 [eBook #72494]

Language: Dutch

Original publication: Amsterdam: Roman- Boek- en Kunsthandel,

1910

Credits: Jeroen Hellingman and the Online Distributed

Proofreading Team at https://www.pgdp.net/ for Project
Gutenberg

*** START OF THE PROJECT GUTENBERG EBOOK LORD LISTER

NO. 0306: EEN AVONTUUR VAN KONING ALFONSO ***
[Inhoud]

[1]

[Inhoud]

☞ Elke aflevering bevat een volledig verhaal. ☜

UITGAVE VAN DEN ROMAN-, BOEK- EN KUNSTHANDEL—SINGEL 236,—
AMSTERDAM.

[Inhoud]
Een Avontuur van Koning Alfonso.
HOOFDSTUK I.
 Het raadsel der schijndooden.

Omstreeks het begin van November bevatte de eerste ochtendeditie

van de „Times” een tamelijk lang bericht, hetwelk in de hoogste mate
de belangstelling der lezers opwekte.

Zie hier de onverkorte inhoud van dit stuk hetwelk onder de

„Gemengde Berichten” was geplaatst:

Het geheim van een Ziekenhuis.

Moordaanslag op een gewonde!

„Het bekende ziekenhuis in de Sloanestreet is, zooals aan onze lezers

bekend zal zijn, reeds sedert enkele dagen het tooneel geweest van
geheimzinnige voorvallen.

In den afgeloopen nacht zijn deze bekroond door een gebeurtenis, die de
gemoederen van alle bewoners van het ziekenhuis, van de directie af tot
aan de minste hulpverpleegster en alle zieken, ten zeerste heeft ontroerd.

Eenige dagen geleden werd er in het ziekenhuis een zwaargewond man

gebracht, die uit verscheidene messteken bloedde en die gevonden was
door twee late voorbijgangers, die opgaven dat zij er getuigen van waren
geweest, hoe de gewonde man overvallen werd door een paar kerels, die
echter op hun nadering de vlucht hadden genomen.

Aanstonds zal blijken, dat er van deze bewering waarschijnlijk geen woord
waar was.

Reeds den volgenden dag werd de zwaargewonde bezocht door een paar
geheimzinnige heeren, die voorgaven tot zijn naaste familie te behooren en
hem dringend wenschten te spreken.

Gedurende het korte gesprek, dat nu volgde, verzocht de gewonde, die

verklaarde Dubois te heeten, aan de hoofdverpleegster der algemeene
mannenzaal, waar hij lag, te telefoneeren naar het logement „Het Vergulde
Hert” en daar navraag te doen naar een zekere Miss Bispham. De
gewonde scheen zeer opgewonden te zijn, het smartte hem [2]blijkbaar
zeer, toen hij ten antwoord kreeg, dat deze dame niet meer in het hotel
aanwezig was.

De beide bezoekers vertrokken, maar werden in de vestibule, juist toen zij

de lift zouden verlaten, gearresteerd door een tweetal particuliere
detectives—bedriegers, zooals later blijken zal!

Want reeds in den loop van denzelfden nacht werden de twee

geheimzinnige bezoekers van Dubois in hetzelfde ziekenhuis binnen
gebracht waar hij ook lag, een hunner zelfs in de groote mannenzaal. De
chauffeur, die hen beiden en de zoogenaamde detectives naar een huis in
het noorden van de stad had gereden, was nieuwsgierig geworden, en op
de loer gaan staan. Hij had er een paar agenten bijgehaald en deze
hadden in het huis twee volkomen bewustelooze personen gevonden, die
naderhand de beide bezoekers bleken te zijn.

Men stond hier voor een raadselachtig geval en de beste geneesheeren,

de directeur van het ziekenhuis dr. Hudson en die geneesheeren
Bushgrave en Greenway begrepen niet het minste van de eigenaardige
bewusteloosheid, waarin de twee mannen verkeerden.

Men kan dus over hun blijdschap en verwondering oordeelen, toen de

befaamde Fransche psychiater Dr. Edouard Dumoulin, zich met het geval
bemoeide en dringend verzocht de beide bewusteloozen te mogen
onderzoeken.

In een vorige editie hebben wij uitvoerig melding gemaakt van deze zaak
en Dr. Dumoulin wist de beide bewusteloozen, die wel schijndood leken
onder zijn wil te brengen, gelastte hen op te staan en zich aan te kleeden
en voerde hen weg. Ongeveer een uur later bleek pas dat de beroemde
Parijsche geneesheer van de geheele zaak niets afwist en verzekerde dat
men zijn plaats had ingenomen met een doel dat hij niet begreep.

Maar nog dienzelfden avond kwam er eenig licht in de zaak, want de beide
mannen, die in hun toestand van volkomen machteloosheid in een groot
restaurant in „Temple Bar” met twee andere heeren dineerden werden op
hun aanwijzing gearresteerd en bleken twee hoogstgevaarlijke bandieten;
bekend onder de bijnamen „Big Billy” en „Bittere Pil”, naar wie de politie
van onze stad reeds geruimen tijd zocht.
Maar toen men wilde omzien naar de beide personen met wie ze waren
binnengetreden en die hen hadden laten arresteeren, waren zij spoorloos
verdwenen!

Wat de beide bandieten betreft, zij wisten zich volstrekt niet meer te
herinneren wat zij gedaan of gezegd hadden van het oogenblik af dat zij
door de gewaande detectives in een huurauto van het gasthuis waren
weggeleid,—naar zij meenden met de gevangenis als bestemming—en
toen ieder een sterke prik in den pols hadden gekregen, waarop bijna
aanstonds het bewustzijn uit hun was weggevloden, althans zij herinnerden
zich volstrekt niets meer, tot op het tijdstip, dat zij tot hun groote
verwondering in de eetzaal van het restaurant in „Temple Bar” weder tot
bewustzijn kwamen.

En toch moeten deze mannen hebben kunnen zien en hooren, zij moeten
zich hebben kunnen bewegen, maar zij deden dit slechts op bevel van een
ander!

De zaak wend intusschen hoe langer hoe geheimzinniger! Wie was de

zwaargewonde in de ziekenzaal, die zulke eigenaardige bezoeken had
gekregen? Wie was de schoone jonge vrouw met het bleeke gelaat, die
voorgaf mevrouw Dubois te heeten, en die hem enkele keeren bezocht,
steeds in gezelschap van een grijsaard met spierwit haar, die haar vader
moest zijn.

De lezers zullen het spoedig vernemen.

In den afgeloopen nacht had er in het ziekenhuis aan de Sloane Street

opnieuw een opzienbarend voorval plaats.

Omstreeks half een kwam zich een jonge verpleegster aanmelden, die
zeide, door mevrouw Dubois gezonden te zijn met het doel om haren
echtgenoot speciaal op te passen, daar zijn toestand, een overbrenging
naar een afzonderlijke kamer, onmogelijk maakte in de eerste dagen.

De hoofdverpleegster stond haar dit verzoek geredelijk toe en het jonge

meisje nam haar plaats aan het ziekbed in, en kweet zich, volgens de
getuigenis [3]van de oude dame, met groote vaardigheid van haar
verpleegsterplichten.

Om één uur werd de groote zaal geheel donker gemaakt, of althans was zij
toen slechts verlicht door de kaarslantaarns van twee surveillanten.
Dezen waren onder den invloed van de stilte van de groote zaal
ingesluimerd, toen zij plotseling werden opgeschrikt door het op luiden toon
gegeven bevel: „Handen op”.

De jonge verpleegster had een der lange linnen gordijnen voor de groote
balkonramen terzijde gerukt, een electrische zaklantaarn opgestoken en
hield nu een revolver gericht op het voorhoofd van een man die achter het
gordijn had gestaan dicht bij het bed van den gewonde Dubois en blijkbaar
zooeven door het balkonraam was binnengedrongen.

De koele luchtstroom die door dit open raam naar binnen drong en die ook
het gordijn zachtjes deed bewegen, had zijn aanwezigheid verraden.

De man werd door de beide surveillanten met behulp van een stuk van het
gordijnkoord gebonden en daarop werden zijn zakken doorzocht.

Deze bleken behalve een revolver en een kleine scherp geslepen dolk, ook
een klein houten étui te bevatten, waarin zich een vaccinatienaald bevond,
waarvan de punt gedrenkt was in een uiterst giftige stof.

De man werd natuurlijk aanstonds in arrest gesteld, maar toen men nu de

verpleegster nader wilde ondervragen, bleek ook zij verdwenen te zijn.

Maar het eigenaardigste komt nog.

Want Big Billy, Bittere Pil, zoowel als de moordenaar achter het gordijn,
inziende dat alles voor hen verloren was, en dat hun een gevangenisstraf
van tientallen jaren wachtte, brachten onder andere aan het licht, dat de
zoogenaamde Dubois niemand anders was dan een Parijsche bandiet, een
man van adellijke afkomst en een authentiek markies Raoul Beaupré de la
Sardogne geheeten.

Hij was ook niet gewond door een overval, maar in een tweegevecht met
een mededinger.

Deze opzienbarende onthulling werd spoedig bevestigd, toen uit Parijs

door de Sureté een uitvoerig signalement werd overgeseind, dat in alles
overeenstemde met het uiterlijk van den gewonde, behalve wat den baard
betreft, dien hij in Engeland waarschijnlijk aanstonds had afgeschoren.

Te Parijs heeft deze markies nog een zestal jaren gevangenisstraf te goed,
en daar hij ook hier in Londen daadwerkelijk aandeel heeft genomen aan
een paar inbraken, zoo kan men wel zeggen dat het een goede dag is voor
onze politie.

Veel werd nu ook opgehelderd. Bijvoorbeeld de identiteit van de

zoogenaamde Madame Dubois, die waarschijnlijk niemand anders was
dan een Fransche bandiete, een zekere Marthe Debussy.

Waar deze vrouw zich echter thans bevindt is niet uit te maken; zeker
logeerde zij niet langer in het „Vergulde Hert”.

En wat het bezoek van de beide naderhand bewusteloos gevonden

schurken van den markies betreft, zij kwamen hem slechts dreigen dat zijn
minnares zich in handen van hun chef bevond en dat deze haar zou
dooden indien Beaupré niet zekere geheimen van zijn eigen bende verried
en zich verbond onder eede, hun chef nooit meer te bestrijden.

Het spreekt vanzelf, dat Dubois, alias markies Beaupré nu aanstonds

onder zeer strenge bewaking is geplaatst en met groote behoedzaamheid
naar een afzonderlijk vertrok is overgebracht, speciaal voor dergelijke
gevallen bestemd, met een zwaar getralied venster, een ijzeren deur en dat
op de vijfde verdieping van het ziekenhuis gelegen is. Zoodra de bandiet
genezen is, zal hij voor zijn rechters moeten verschijnen, en daarna zijn
rechtvaardige straf ondergaan.

De man achter het gordijn, eveneens een berucht misdadiger, trachtte zijn
straf blijkbaar verminderd te krijgen door te verklaren, dat hij handelde in
opdracht van den chef zijner bende, wiens naam hij echter weigerde te
noemen, evenals trouwens „Big Billy” en „Bittere Pil”.

Hij was gekomen om Beaupré te vermoorden, en aldus den aanvoerder te

beschermen tegen zijn aanval.

En voor de rest wilde hij niets loslaten. [4]

Met dit al moeten er nog vrij wat raadsels door de politie worden opgelost.

Zoo weet zij bijvoorbeeld volstrekt niet, wie de „edele grijsaard” was, die als
vader van de zoogenaamde mevrouw Dubois optrad, noch wie de beide
detectives waren, noch wie de rol van dr. Dumoulin vervuld heeft, en
evenmin wie op zulke geniale wijze de taak van verpleegster heeft vervuld,
maar de politie heeft vermoedens. Er wordt een naam gefluisterd, dien wij
nu niet zullen herhalen, ten einde het onderzoek van de politie niet te
 bemoeilijken, maar die waarschijnlijk dezer dagen alom bekend zal zijn
gemaakt; doch als Scotland Yard er ook ditmaal niet in slaagt hem
gevangen te nemen, dan zal het wel eeuwig een raadsel blijven wat hem er
toe bewoog als beschermer op te treden voor een man als Beaupré, daar
deze zich in een halsstarrig stilzwijgen hult, en weigert eenige inlichtingen
te geven.

In ieder geval is het niet te loochenen, dat de politie, dank zij het ingrijpen
van den geheimzinnigen man, wiens naam thans nog niet genoemd wordt,
wederom vier zeer gevaarlijke bandieten in handen heeft gekregen.

[Inhoud]
HOOFDSTUK II.
 Belofte maakt schuld.

Dit bericht in de „Times” verwekte, zooals gezegd, niet weinig

opschudding, niet alleen in kringen der politie, maar onder de
geheele burgerij, die het geval dagen lang besprak, terwijl men
daarbij de zonderlingste meeningen hoorde.

Onder die burgerij mag ook gerekend worden John Raffles, de

Groote Onbekende, die in dit alles zulk een groote rol had gespeeld,
met behulp van zijn trouwen vriend Charly Brand en den braven reus
Henderson, zijn chauffeur.

De gentleman-inbreker las het stukje, terwijl hij gezeten was in de vrij

schaars gemeubelde ontvangkamer van een splinternieuw huis in
een der noordelijkste wijken van Londen, gelegen in een straat,
waarvan de aanleg nog niet lang geleden begonnen was.

Zijn hoed en zijn stok lagen naast hem, en hij was blijkbaar zooeven
pas gekomen.

Het was tien uur in den morgen, toen Raffles eenig gerucht in een
aangrenzend vertrek hoorde, en het volgende oogenblik ging de deur
open en trad er een jonge vrouw met een mooi, maar zeer bleek
gelaat, binnen, die haastig op Raffles toetrad en zeide:

—Ik vraag u verschooning als ik u heb laten wachten, mijnheer—ik

heb een zeer slechten nacht doorgebracht!

—Het is eerder aan mij om u mijn verontschuldigingen aan te bieden,

dat ik u op zulk een onbehoorlijk [5]uur kom bezoeken, Madame,
hernam Raffles, die was opgestaan.

Er lag een ernstige klank in zijn stem toen hij vervolgde:

—Gij zult begrijpen, dat ik hier op dit uur niet zou gekomen zijn,
Marthe Debussy, wanneer het geen ernstige aangelegenheid gold, gij
moet sterk zijn—er is met uw vriend, met Raoul Beaupré, iets
voorgevallen hetwelk ik wel voorzien had, maar toch onmogelijk kon
vermijden.

De jonge vrouw wankelde en moest zich aan den rand van de tafel
vastgrijpen, terwijl zij de linkerhand op het hart drukte.

Bijna toonloos vroeg zij:

—De aanslag heeft dus plaats gehad en—wat?

—Stel u gerust, mijn vriend heeft den aanslag zelf kunnen verijdelen,
maar helaas niet de gevolgen! Big Billy zoowel als de Bittere Pil en
de man die den aanslag had willen plegen, hebben uw vriend
verraden, en zijn waren naam aan de politie opgegeven. Hier—lees.

Hij stak de jonge vrouw het nummer van de „Times” toe, en zij liet
zich op een stoel neervallen, nam het blad met een automatische
beweging aan, en trachtte te lezen.

Maar reeds na de eerste regels gaf zij Raffles het blad terug en zeide
met bevende stem:

—Ik kan niet! Het warrelt mij voor de oogen. Ik smeek u! lees gij het
mij voor.

Raffles nam nu het blad en las met heldere stem het bericht voor.

De jonge vrouw had al dien tijd bewegingloos zitten luisteren, de

zwarte oogen op den zwarten vloer gevestigd, de slanke witte vingers
ineen gestrengeld—het toonbeeld der wanhoop!
Toen Raffles de voorlezing van het noodlottige bericht had beëindigd,
bleef het geruimen tijd stil in het vertrek.

In dit huis, hetwelk hij slechts een week te voren gehuurd had, had
Raffles de vrouw, wie hij zijn hulp had toegezegd, omdat zij door zijn
toedoen in dezen toestand was gebracht en in groot gevaar had
verkeerd, voorloopig ondergebracht, ten einde haar buiten den greep
van den langen arm der justitie te houden.

Zij mocht volstrekt niet uitgaan zonder zijn toestemming en werd daar
steeds bewaakt, door Raffles zelf, door Charly Brand of door
Henderson, omdat haar leven gevaar liep. Zij immers had aan Raffles
het plan voor een inbraak verraden, ten einde Dr. Fox, den chef van
een groot misdadigersgenootschap, en de persoonlijke vijand van
haren minnaar, in handen der politie over te leveren en zich aldus te
wreken over de zware wonden, welke hij Beaupré had toegebracht.

Zij had zich reeds in handen van de mannen van Dr. Fox bevonden,
en het doodvonnis zou zeker aan haar voltrokken zijn geworden, als
Raffles en zijn twee vrienden niet tusschen beide waren gekomen.

De Groote Onbekende was de eerste die weder sprak. Hij had

eenigen tijd ernstig nagedacht en zeide nu:

—Ik mag er ook tegenover u geen geheim van maken Madame.

Raoul Beaupré is en blijft mijn vijand! Misschien zal de politie ons
over een kam scheren—ik denk daar echter anders over! Ja—ik weet
wat gij zeggen wilt: Ik heb mij niet als een vijand gedragen! Vergeet
echter niet dat ik dit voor een deel uit eigen belang deed! Als ik de
vijanden van Dr. Fox help, dan bestrijd ik hem zelf, maar er was nog
een overweging, die alle anderen in de schaduw stelde.

—Wat meent gij, vroeg Marthe Debussy op zachten toon.

—Gij hadt mijn woord! Het was mijn schuld dat gij in groot
levensgevaar hebt verkeerd en ik heb beloofd, u te beschermen en
wat uw vriend betreft—hij is in zekeren zin mijn bondgenoot, en wel
een zeer kostbare, want hij is een doodsvijand van Fox! en ik zou hier
met een kleinen variant kunnen zeggen: „De vijanden van mijn
vijanden zijn mijn vrienden”.

—Gij wilt dus zeggen …? stamelde Marthe, terwijl haar oogen vochtig
begonnen te glanzen.

—Ik wil zeggen, dat ik mijn hand nog niet van hem aftrek, maar dat ik
wil trachten hem toch nog voor u te redden.

De jonge vrouw liet een snikkend geluid hooren, greep de hand van
Raffles en voor hij het had kunnen verhinderen, had zij er een kus op
gedrukt.

—Dat moet gij niet doen, zeide Raffles hoofdschuddend, terwijl hij
zijn hand snel terug trok. Ik zeg u immers, dat ik handel uit eigen
belang! Ik wil u echter niet ontveinzen, dat het ditmaal zeer moeilijk
zal zijn. Ik vrees zelfs dat het volkomen onuitvoerbaar zal zijn,
zoolang Beaupré in het ziekenhuis ligt! Wij moeten [6]echter alles in
het werk stellen, om hem nog te redden vóór hij naar de gevangenis
wordt overgebracht, want daar zal het wellicht onmogelijk zijn.

—Kan ik u daarbij niet van dienst zijn, vroeg Marthe Debussy op

zachten toon. Ik zou u zoo gaarne helpen! Ik heb niets anders ter
wereld dan Raoul, en als ik mijn leven voor hem zou moeten offeren,
dan zou ik geen seconde twijfelen.

Raffles had even nagedacht, en zeide toen:

—Misschien kunnen wij uw hulp gebruiken, dan zal ik u wel laten

weten wat wij in de eerste dagen kunnen doen, want bij mijn
pogingen om Beaupré te bevrijden zal ik ook zijn eigen hulp noodig
hebben; hij moet mede werken, en daartoe is hij thans nog veel te
zwak. Maar ik beschik over middelen, waardoor ik de natuur ter hulp
zal kunnen komen en zijn herstel zal kunnen bespoedigen. De zaak
is slechts of het mogelijk zal zijn, tot hem door te dringen om hem dat
middel toe te dienen, want de directie zal nu natuurlijk zeer
wantrouwend zijn en goed acht geven op ieder die binnentreedt. Ik
vrees dat wij den truc met de zoogenaamde verpleegster niet
nogmaals kunnen toepassen—en om voor de tweede maal als dokter
Dumoulin op te treden, dat zal ook wel niet gaan. Wij hebben echter
den tijd om daarover na te denken, want in normale omstandigheden
zou het zeker nog minstens een week duren alvorens men Beaupré
naar de gevangenis zou kunnen overbrengen.

Raffles was opgestaan en had hoed en stok gegrepen, hij maakte

een buiging voor de jonge vrouw en zeide:

—Ik kom u vanmiddag halen om u met een mijner auto’s naar een
andere stad te brengen, hier dicht bij, want hier zijt ge nog altijd niet
veilig genoeg, en wij zouden u ook niet goed meer kunnen bewaken,
daar ik bij mijn plannen ter bevrijding van uw vriend de hulp van mijn
twee makkers niet zou kunnen ontberen. Als gij uw haar verft, of een
pruik opzet, zult gij vrijwel veilig zijn in een of andere provinciestad.
Kleed u zoo eenvoudig mogelijk en vertoon u tijdelijk zoo weinig
mogelijk in het publiek. Indien wij uw hulp kunnen gebruiker, zullen wij
u aanstonds laten halen!

—Ik voeg mij geheel maar uw aanwijzingen, mijnheer zeide Marthe

Debussy op zachten toon. Hoe laat komt gij?

—Om drie uur!

Nogmaals boog Raffles voor de jonge vrouw en het volgende

oogenblik had hij het vertrek en het huis verlaten.
Honderd schreden verder stapte hij in een wachtende auto, die hem
naar het einde van de Regent Street bracht, dicht bij Pall Mall.

In de eerstgenoemde straat bewoonde Raffles, onder den naam van

Lord William Aberdeen, een fraai heerenhuis, dat met een voor deze
buurt zeer grooten tuin omgeven was.

Hij kon dit huis steeds en ten allen tijde ongezien binnengaan, in
welke vermomming hij zich ook bevond, want een der tuinmuren
strekte zich over geruimen afstand uit langs een straat, welke zoo
goed als nimmer begaan werd, en welker overzijde werd ingenomen
door den blinden muur eener opslagplaats, tijdens den oorlog door de
regeering overgenomen en thans in het bezit van Raffles zelf, die het
groote gebouw voordeelig had verhuurd, op voorwaarden dat er aan
de bestemming niets mocht worden gewijzigd.

Niemand bespeurde er dus iets van, dat hij de kleine deur in den
tuinmuur opende en snel binnentrad.

Hij stak den grooten tuin dwars over en ging het huis aan de
achterzijde binnen.

Daarop beklom hij een smalle trap, die hem rechtstreeks naar zijn
slaapkamer bracht, waar hij de vermomming kon afleggen, welke hij
gedragen had bij zijn bezoek aan de minnares van den Franschen
Markies.

Terwijl hij hier mede bezig was, werd er op de deur geklopt en de

stem van Charly Brand liet zich hooren.

—Ben je hier, Edward?

—Ja, Charly! Wacht—ik zal open doen!

Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

textbookfull.com