Microprocessors and Microsystems 91 (2022) 104514

Contents lists available at ScienceDirect

Microprocessors and Microsystems

journal homepage: www.elsevier.com/locate/micpro

A hardware framework for fall detection using inertial sensors and

compressed sensing
Oussama Kerdjidj a ,∗, Elhocine Boutellaa a , Abbes Amira b,c , Khalida Ghanem a , Fatima Chouireb d
a
Center for Development of Advanced Technologies, Algiers, Algeria
b
Department of Computer Science, University of Sharjah, Sharjah, United Arab Emirates
c
Institute of Artificial Intelligence, De Montfort University, United Kingdom
d
Laboratory of Telecommunications, University Amar Telidji, Laghouat, Algeria

ARTICLE INFO ABSTRACT

Keywords: Falls are the leading cause of sudden and accidental injuries in the elderly. With the aging of the world’s
Fall detection population, one concern of the elderly increase as well. Notably, the rate of fall accidents and hospitalizations
Human activity related to falls is increasing year after year. Monitoring human activity to detect falls using intelligent
Wearable sensors
portable systems is an efficient and economical solution that enables faster intervention and immediate
Compressed sensing
action. This article proposes a hardware framework for fall detection using accelerometer and gyroscope data
Classification
implemented on a Zedboard FPGA (Field Programmable Gate Array). Hardware components are designed,
tested and simulated using the Xilinx Vivado tool. On this proposed design, the innovation lies in the insertion
and integration of the compressed sensing technique (CS) in the fall detection system and human activities
to reduce a number of samples and thus reduce energy consumption. To this end, four hardware blocks
(compression, recovery, feature extraction and prediction) are designed and tested and validated against the
software implementation. In the same system, we implemented the design by exploiting pipeline optimization,
achieving very low latency compared to the unoptimized configuration.

1. Introduction the research community before and during the COVID-19 pandemic.
This attention is reflected by the number of published works addressing
Ageing people are the most vulnerable category in current society the fall detection and activity recognition challenges. Depending on
due to many factors, such as chronic diseases, Osteoporosis, difficulty the sensing technology used, these works can be categorized into
to move, loneliness, etc. These factors affect older adults and increases environment sensor based [2] and wearable sensor based [3–6]. In the
the risk of falls during their daily living. Several critical damages can
first category, we find vision systems based on cameras [7–9] as well
affect a person’s health after a fall, and sometimes they become fatal.
as other sensors such as microphones [10], Radio-frequency identifi-
Furthermore, even though the risk of falls is a big concern for elderly
people, it incurs other categories, such as pregnant women, children, cation (RFID) Tags [11], Infrared sensors (IR) [12], etc. In the second
etc. According to the World Health Organization (WHO) [1], the num- category, we find accelerometers, gyroscopes, magnetometer [13–15],
ber of fall accidents is increasing year after year. On the other hand, the etc. We are interested to the second category due to some inherent
pandemic-related restrictions can instigate a vicious cycle of reduced advantages like wide area of coverage (indoor and outdoor, practically
physical activity. Particularly, old persons have a higher probability everywhere), low cost and ease of deployment.
of contracting severe complications from coronavirus (COVID-19). The Wearable fall detection enabled devices are embedded systems in-
particular recommendation of the elderly is to follow strict guidelines tended to be carried on the body most of the time. Most of the literature
to minimize the risk of exposure to contamination during the current works [16–18] focus on the development and validation of the software
pandemic. Therefore, designing and developing a system by exploiting part of the fall detection system comparing to the hardware part. How-
the new technology to monitor the physical activity and fall events ever, a co-design approach is more appropriate for the development of
remotely reduces direct contact with these vulnerable adults and thus
such a system because it takes into consideration targeted resources.
decreases the risk of their contamination.
Therefore, in this work we give the hardware implementation the
Considering its importance and impact on society, fall detection and
human activity monitoring have attracted significant attention from importance it deserves. Since Field Programmable Gate Array (FPGA) is

∗ Corresponding author.
E-mail address: [email protected] (O. Kerdjidj).

https://doi.org/10.1016/j.micpro.2022.104514
Received 27 October 2021; Received in revised form 1 February 2022; Accepted 28 March 2022
Available online 6 April 2022
0141-9331/© 2022 Elsevier B.V. All rights reserved.
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

Fig. 1. Overview of the proposed fall detection system.

a prototyping platform, we adopt it for the design and implementation and each one performs a specific task. The Compression and Recovery
in our work. Additionally, our work exploits compressed sensing tech- modules, respectively, ensure signal compression and recovery based
nique as it allows hardware resource optimization while guaranteeing on the CS technique. The Feature module extracts discriminate informa-
acceptable system performances. tion from the signal. Finally, the k-NN module classifies the extracted
A number of previous works explored the compressive sensing characteristics to detect falls and recognize human activities.
advantages for recognizing daily activities and detecting falls based on The organization of the paper is as follows. Section 2 presents
wearable sensors at a software level. For instance, Kerdjidj et al. [19] the proposed fall and daily activity detection hardware design and
applied CS to the inertial sensors aiming to reduce the data transmission implementation exploiting the compressed sensing technique. In Sec-
and energy consumption. The body inertial signals acquired with a tion 3, we evaluate the results of the resource occupation for various
Shimmer platform are compressed before being transmitted to a base modules of our system and provide the global system design on the
station where these signals are recovered and processed to perform fall targeted platform. Section 4 summarizes the paper and traces further
detection and activity recognition. In another work, Gibson et al. [20] improvements.
applied the wavelet transform to the signals accelerometer collected
by the Shimmer platform. A thresholding function and a pre-known 2. Proposed approach
random index permutation are exploited to improve the sparse signal.
The Shimmer device executes the compression before transmitting data. Fig. 1 illustrates the global overview of our proposed system on the
Once received, the signal is recovered and processed to detect the falls. Zynq circuit [32]. Our proposed design of fall detection and human ac-
Four reconstruction algorithms are studied: stagewise OMP (StOMP), tivity recognition using compressed sensing is divided into four blocks,
regularized orthogonal matching pursuit (ROMP), orthogonal matching compression, recovery, feature extraction and classification. One major
pursuit (OMP) and matching pursuit (MP). The classification using Prin- aspect of the proposed approach is the exploitation of the compressed
cipal Component Analysis (PCA) and decision tree is performed in the sensing technique to reduce the energy consumption by decreasing the
discrete wavelet transform (DWT) domain avoiding the application of number of samples to transmit for an eventual reconstruction of the
the inverse transform. In Djelouat et al. [21] the acquired acceleration signal to predict a normal movement or a fall event. The CS technique
data by means of Shimmer devices is compressed via multiplication ensures that the original signal can be reconstructed from a few samples
by a binary sensing matrix and recovered using subspace pursuit. Two under some conditions. Our system utilizes a KNN classifier powered by
classification approaches were investigated: k-nearest neighbor (KNN) features extracted from the reconstructed signal using the compressed
and extended nearest neighbor (ENN). When it comes to hardware, sensing technique. All the four modules are implemented on a hardware
only few works consider the CS implementation [22–29]. For example, platform as an embedded system to accelerate the computations. The
Kerdjidj et al. [22] designed a module to recover a signal deployed in body inertial data is captured with a Shimmer device [33], which
healthcare by exploiting the compressed sensing technique on the Zynq provides gyroscope, accelerometer, magnetometer signals, as well as,
circuit. In an other work, Djalout et al. [23] implemented and exploited physiological signals such as Electromyogram (EMG) and Electrocardio-
the CS technique on Zynq SoC to transmit large data with high privacy gram (ECG). Though physiologic signals provide important information
and identify the patient using the ECG signal. about the body, we limit our present work to the kinematic signals in
In terms of hardware based classification, the k-Nearest Neighbors this study.
is the most widely implemented classifier due to its simplicity and effi- The challenge in our paper is the hardware implementation of
ciency [30,31]. For instance, Xiaojia et al. [30] used the Vivado High- the whole system, but a software implementation on Matlab is more
Level Synthesis (HLS) to decrease the influence of the memory-access than necessary to extract the best parameter of various blocks. So,
constraint in KNN. To this end, the authors employ two data access after several tests, validations, and optimization of the approach at
reduction methods, namely, PCA based filtering and low-precision data the software level. The experimental evaluation has shown promising
representation. results, achieving 93% accuracy, where the system recognizes seven
Most of the reviewed works address the problem at a software classes that include the fall event along with six activities of daily living
level or implement a part of the system. In the proposed work, the (standing, running, walking, jumping, . . . ). The software validation
design and the implementation of the whole system on the same allowed setting the best system parameters. Indeed, we found a window
platform is taken on considered. Our design consists of four modules, length of 128 timestamps and a dictionary of size 128 × 100 to perform

2
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

Algorithm 1 Matching Pursuit

Input: 𝛷: Sensing matrix; Y: Observation vector; K: Sparsity of signal X.
Output: X̂: Original signal estimation
Procedure:
Initialize residual and set iteration counter
𝑟0 = 𝑌
𝑖=1
Find index
𝜆𝑖 = argmax |⟨𝑟𝑖−1 , 𝜙𝑗 ⟩| for 𝑗 = 1 … 𝑁.
Calculate the new approximation and the new residual
𝑋̂ 𝑖 = 𝑋̂ 𝑖−1 +⟨𝜙𝜆𝑖 , 𝑟𝑖−1 ⟩𝜙𝜆𝑖
𝑟𝑖 = 𝑌 − 𝑋̂ 𝑖−1 = 𝑟𝑖 − ⟨𝜙𝜆𝑖 , 𝑟𝑖−1 ⟩𝜙𝜆𝑖
Increment iteration counter 𝑖 = 𝑖 + 1,
return to step 2 if 𝑖 < 𝐾
Output estimation 𝑋̂

the best. The compression dictionary should be pre-computed; we 2.1.2. Recovery block
directly deploy the software-generated dictionary on the hardware. In The second Intellectual Property (IP) role is to recover a compressed
the next section, we provide a detailed description of each building signal from few samples compacted previously in the compress block.
block of our system. As input, this block requires the signal coefficients and the dictionary
indexes computed by the compressed block. Given the compression dic-
2.1. Compressed sensing tionaries, its components corresponding to the indexes are multiplied
by the coefficients and summed to yield the original signal. Fig. 2-b
In this section we explain how we employ CS and the used algo- illustrated the diagram of the proposed design of the recovery block.
rithm for hardware based fall detection using accelerometer. The CS
algorithm compresses the data collected by the Shimmer platform re-
duces the transmission cost. Once received, the compressed signals are Algorithm 3 Recovery
transmitted and processed to cover the original data. The CS technique Input: Index and Coef
is able to recover the signal using few samples satisfying sparsity and Output: 𝐵𝑟𝑒𝑐 Signal recovery
incoherence conditions [34,35]. Download the signal coefficients (𝐶𝑜𝑒𝑓 ) and indexes (𝐼𝑛𝑑)
In our work, we utilize the Matching Pursuit (MP) algorithm for 𝐼𝑛𝑑1, … 𝑖𝑡𝑒𝑟 = 𝑖𝑛𝑑𝑒𝑥1 , 𝑖𝑛𝑑𝑒𝑥2 , … 𝑖𝑛𝑑𝑒𝑥𝑖𝑡𝑒𝑟
signal compression and recovery. It is from a greedy family algorithm. 𝐶𝑜𝑒𝑓1, … 𝑖𝑡𝑒𝑟 = 𝑐𝑜𝑒𝑓1 , 𝑐𝑜𝑒𝑓2 , … 𝑐𝑜𝑒𝑓𝑖𝑡𝑒𝑟 (Pipleline)
MP algorithm ensures a compromise between computation complexity Extract the data from dictionary A
and compression performance compared to other methods [36]. Its NewMat1∶𝐼𝑛𝑑 = A1 , A2 , … A𝐼𝑛𝑑 (Pipleline)
intuition is to approximate the signal by a weighed sum of few columns Reconstruct the signal B𝑟𝑒𝑐
forming the representation dictionary. MP is a greedy solution for the B𝑅𝑒𝑐 = NewMat1∶𝐼𝑛𝑑 × 𝐶𝑜𝑒𝑓1, … 𝑖𝑡𝑒𝑟 (Pipleline)
problem of searching the sparsest representation of a signal. This is
achieved by iteratively finding the best projection of the data onto the
span of the coding dictionary, which sparsely approximates the signal.
The Pseudo-code 1 illustrated the MP algorithm. 2.2. Features extraction

2.1.1. Compressed block The input of this block is the data of the accelerometer and the
Fig. 2-a illustrates the first part of the MP algorithm. This algorithm gyroscope signals reconstructed by the recovery block. As output,
enjoys more straightforward hardware implementation compared to the the block generates a feature vector. To ensure real-time operation
OMP variant while MP can perform as OMP with a specific design on of the proposed fall detection system, we selected statistical features
hardware. The main goal of this architecture consists in compressing with reduced complexity. This ensured a high computational speed
the device acquired data (accelerometer and gyroscope), we assume and easy implementation of the feature extraction block. Table 1 il-
the high rate compression in this work compared to additional per- lustrates the equations used in our work to extract the parameters
formance [22,37]. The compression block inputs are the dictionary from the accelerometer signals (𝑥, 𝑦 and 𝑧). The selected features are
and the signal to be compressed. The block produces as outputs the the maximum, the minimum, the average, the number of peaks, the
coefficients and the indexes after high computation.
standard deviation, the Root mean square of each accelerometer axis.
Additionally, we compute the Pearson correlation between pairs of the
Algorithm 2 Compression three accelerometer signal axes (X–Y, X–Z and Y–Z). These parameters
Input: 𝐴: Sensing matrix; B: Signal are computed from windowed accelerometer signal of length 2.56 s,
Output: Index and Coef corresponding to 128 samples, since the sampling rate of the device
Compute the correlation (𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛) is 50 Hz. In total, we obtained forty-five features computed by the
- 𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 = A × B (Pipeline) corresponding module for both signals (accelerometer and gyroscope).
Find the index (𝐼𝑛𝑑𝑒𝑥) of the best correlation
- Index = max(𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛) 2.3. Classification
Compute the Coefficients (Coefficient)
- Coefficient = A𝐼𝑛𝑑𝑒𝑥 × B (Pipeline) The last part of our implementation involves using a machine learn-
Update the signal B ing technique to classify the data (daily activity) and detect the falls.
- 𝐵𝑛𝑒𝑤 = 𝐵𝑜𝑙𝑑 - [𝐴𝑖𝑛𝑑𝑒𝑥 ×Coefficient] (Pipeline) To this end, we employ the k-NN model due to the simplicity of im-
plementation and its satisfactory accuracy compared to more complex
algorithms. k-NN computes the distance between the features vectors

3
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

Fig. 2. Compress (a) and recovery (b) block diagrams.

Table 1
Statistical equations used to extract features. The implementation of the k-NN algorithm on Vivado HLS, requires
Feature Equation storing the reference feature vectors on the platform, the computation
- Maximum max(𝑥𝑖 ) of the query distance to each reference as well as sorting the result to
- Minimum min(𝑥𝑖 )
1 ∑𝑁
identify the nearest feature.
- Mean 𝑁 𝑖=1 𝑥𝑖
- Number of Peaks Peaks in which the amplitude is
superior to the average 3. Results and discussion
√
1 ∑𝑁
- Standard Deviation 𝑁−1
(𝑥 − 𝜇)2
𝑖=1 𝑖
√ ∑ This section presents the simulations and results of the four IPs
1 𝑁 2
- Root Mean Square 𝑖=1 (𝑥𝑖 )
𝑁
∑𝑁
(compress, recovery, Features and algorithm of classification k-NN)
𝑖=1 (𝑥𝑖 −𝑥)(𝑦𝑖 −𝑦)
- Pearson correlation √∑𝑁 √∑𝑁 implemented in Vivado-HLS and Vivado. All IPs are independent of
2 2
𝑖=1 (𝑥𝑖 −𝑥) 𝑖=1 (𝑦𝑖 −𝑦)
others; they start working when the previous IP finishes its execution
where 𝑥 and 𝑦 is the mean of signal 𝑥 and 𝑦,
and provides the necessary results for the next one. Table 2 summarizes
𝑁 the length of the signal.
the performance for each IP separately in terms of many parameters.
According to the kind of the synthesis used for the IP, the result
represents the resource occupation area and the latency time for each
and assigns the next sample to one class among the k closest training module. The surface occupation is expressed in terms of block memory
samples. The pseudo-code of the k-NN implementation is depicted in B-ram, Digital Signal Processing DSP, registers FF, and Look Up of
the Algorithm 4. Table LUT. All these parameters are computed either using pipelines
or without any optimization, where the measured Uncertainty is 1.5
Algorithm 4 KNN for all directives. This optimization makes the IP faster but with more
resources.
1. Load the training and test data
The first IP is the Compress block; as mentioned previously, this
2. Choose the value of K = 1
module’s primary goal is to compact and reduce the size of samples into
3. For each point in test data:
smaller vectors. The compress block stores the compression dictionary,
- find the Euclidean distance to all training data points
which takes an essential part of the block’s resource. This explains the
- store the Euclidean distances in a list and sort it
high percentage of B-ram (13%) in Table 2, corresponding to 39 B-
- choose the first k points
ram blocks from the total available B-ram in the device (280). For the
- assign a class to the test point based on the majority of
other parameters, DSP, FF, and LUT are changed slightly according to
classes present in the chosen points
the optimization, but in one case, the occupation is too high due to
4. End
the high number of directives (Num Direc). This significant number of

4
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

Table 2
Latency performance and area occupation.
IP Num Direc Bram DSP FF LUT Latency (Clock cycles) Pipeline
0 1505 (1%) 2338 (4%) 1674292 No
1 2302 (4%) 332573
1497 (1%)
Compress 2 39 (13%) 5 (2%) 2307 (4%) 3313303
Yes
3 1524 (1%) 2357 (4%) 330287
4 1570 (1%) 2523 (4%) 133179
5 10840 (10%) 11017 (20%) 36154
0 892 (0%) 1224 (2%) 15649 No
1 928 (0%) 1334 (2%) 7719
Recovery 42 (15%) 5 (2%)
2 1747 (1%) 1731 (3%) 3922
Yes
3 1732 (1%) 1770 (3%) 2614
4 1876 (1%) 9717 (18%) 2614
0 3575 (3%) 5879 (11%) 3999 No
Features 2 (0%) 19 (8%)
1 3502 (3%) 5891 (11%) 1966
Yes
2 3435 (3%) 5757 (10%) 1203
Predict 0 120 (42%) 16 (7%) 2633 (2%) 4505 (8%) 490253 No
Resource board – 280 220 106400 53200 – –

Table 3
Comparing two configurations of the proposed fall detection system according to the optimization.
IP Num Direc Bram DSP FF LUT Latency (Clock cycles) Pipeline
No optimization 0 203 (70%) 45 (19%) 8605 (6%) 13946 (25%) 2184193 No
With optimization 11 203 (70%) 45 (19%) 18784 (16%) 30996 (56%) 530224 Yes
Available resources – 280 220 106400 53200 – –

Fig. 3. Overview of the four IPs implemented in Vivado.

occupation space of Directive 5 is awe-inspiring, but the reduction of The Third IP is the Features module, which computes the statistical
latency execution appeared (A slowness by 46X time when the IP is values from the three axes of each sensor (accelerometer and gyro-
without any optimization). scope) after being recovered by the previous IP. Compared to the other
The second IP shown in the table is the recovery block. This module IPs, the results illustrate that this IP requires few B-ram resources, but
reconstructs both accelerometer and gyroscope signals to recover a new the used DSP and LUT resources are considerable. This IP exploits (8%)
signal which looks, as match as possible, like the original signal. To of DSP and (10%–11%) of LUT. For the Latency, we notice that the
this end, the same dictionary used in the Compress IP is used with number of execution cycles varies between 3999 and 1203 according to
the compression coefficients and index. As it can be noticed from the the used Directives. The IP will be three times faster when two pipelines
table a total of (15%) B-ram is occupied by this IP corresponding to are used.
42 blocks. While the used device B-ram is significant, few DSP, FF The last IP is the Predict module, which implements the KNN
and LUT resources are required, except when we adopt four directives algorithm. This module receives the feature vector extracted by the
corresponding to four different pipelines inside the IP. We can notice previous IP and computes its Euclidean distance to all the stored
that using four pipelines, the latency of the IP decreases from 15649 to reference features. The output of this IP is the class of the nearest
2614 which is at least 5X times optimization. reference feature. A large number of B-ram blocks used by this IP due to

5
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

need of storing the training features. This IP uses 120 memory blocks Declaration of competing interest
out of 280 on the device (42%), making it the block that consumes
the most memory in our design. The rest of performance is between The authors declare that they have no known competing finan-
(8%) of LUT, (2%) of FF and (7%) of DSP due to arithmetic operations cial interests or personal relationships that could have appeared to
required to compute the Euclidean distance and the distance sorting influence the work reported in this paper.
process. The latency is also very high due to the extensive computations
between features and the sorting process. We were not able to employ Acknowledgments
any optimization in this module because the IP synthesis process is very
The authors thank the ‘Direction generale de la recherche sci-
low (we run it for more than a week without getting any result).
entifique et du développement technologique’ and the ‘ministère de
After designing and testing each IP individually, we integrated the
l’enseignement superieure et de la recherche scientifique’ for funding
four modules on the board simultaneously. At this stage, we considered
this work as part of the project CRCC:N04-2/DTELECOM/CDTA/PT19-
two different configurations of our proposed fall detection system,
21.
the first is without any optimization while the second uses the best
optimization for each IP. The resources of these two architectures are References
compared in Table 3. This optimized configuration employs eleven
directives all IP comprised. With this architecture, latency performance [1] World Health Organization, Falls, 2021, URL https://www.who.int/news-room/
can reach an enhancement of four times compared to the basic one. fact-sheets/detail/falls.
[2] J. Clemente, F. Li, M. Valero, W. Song, Smart seismic sensing for indoor fall
However, the optimized architecture exhibits considerable resources
detection, location, and notification, IEEE J. Biomed. Health Inf. 24 (2) (2020)
occupation. Specifically, (10%) of FF and (31%) of LUT are required. 524–532, http://dx.doi.org/10.1109/JBHI.2019.2907498.
The required B-ram and DSP resources are the same for both ar- [3] X. Chen, S. Jiang, B. Lo, Subject-independent slow fall detection with wearable
chitectures, since no more optimization can be achieved on these sensors via deep learning, in: 2020 IEEE Sensors, 2020, pp. 1–4, http://dx.doi.
org/10.1109/SENSORS47125.2020.9278625.
resources.
[4] T. Nguyen Gia, V.K. Sarker, I. Tcarenko, A.M. Rahmani, T. Westerlund, P.
After creating all the necessary IPs of our system on Vivado HLS, we Liljeberg, H. Tenhunen, Energy efficient wearable sensor node for IoT-based
used the Vivado tool to build a suitable design by exporting the created fall detection systems, Microprocess. Microsyst. 56 (2018) 34–46, http://
IPs to the Vivado library. Fig. 3 illustrates our fall detection scheme dx.doi.org/10.1016/j.micpro.2017.10.014, URL https://www.sciencedirect.com/
science/article/pii/S0141933117301461.
exploiting the compressed sensing technique. The principal elements [5] T. Xu, H. Se, J. Liu, A fusion fall detection algorithm combining threshold-
that constitute the system are the previous IPs (Compress, Recovery, based method and convolutional neural network, Microprocess. Microsyst. 82
Feat and predict). To manage these IPs, the Processing System Reset (2021) 103828, http://dx.doi.org/10.1016/j.micpro.2021.103828, URL https://
provides a mechanism to handle the reset conditions for the system; www.sciencedirect.com/science/article/pii/S0141933121000089.
[6] W. Saadeh, S.A. Butt, M.A.B. Altaf, A patient-specific single sensor IoT-based
the Concat IP is employed to join the various interrupt signals provided wearable fall prediction and detection system, IEEE Trans. Neural Syst. Rehabil.
from the different blocks; and the AXI4 Interconnect link between the Eng. 27 (5) (2019) 995–1003, http://dx.doi.org/10.1109/TNSRE.2019.2911602.
register and the memory. The software is running on Zynq7 Processing [7] O. Keskes, R. Noumeir, Vision-based fall detection using ST-GCN, IEEE Access 9
(2021) 28224–28236, http://dx.doi.org/10.1109/ACCESS.2021.3058219.
System managed directly by the physical connection of DDR memory.
[8] R. Gupta, P. Anand, S. Chaudhury, B. Lall, S. Singh, Compressive sensing based
privacy for fall detection, 2020, arXiv:2001.03463.
4. Conclusion [9] V. Angayarkanni, V. Akshaya, S. Radha, Design of a compressive sensing based
fall detection system for elderly using WSN, Wirel. Pers. Commun. 98 (1) (2018)
421–437, http://dx.doi.org/10.1007/s11277-017-4876-x.
This paper presents an advantageous and interesting system for [10] J.S.R. Alex, M. Abai Kumar, D.V. Swathy, Deep learning approaches for fall de-
human activity recognition and fall detection integrating compressed tection using acoustic information, in: N. Zhou, S. Hemamalini (Eds.), Advances
in Smart Grid Technology, Springer, Singapore, 2021, pp. 479–488.
sensing technique on hardware implementation. The design can rec-
[11] H. Kamoi, K. Toyoda, T. Ohtsuki, Fall detection using UHF passive RFID based on
ognize the daily activities (walk, jump, run, lay, stand, sit) and detect the neighborhood preservation principle, in: 2018 IEEE Global Communications
the fall event by exploiting the machine learning to distinguish seven Conference, GLOBECOM, 2018, pp. 1–6, http://dx.doi.org/10.1109/GLOCOM.
classes. In the same approach, we integrate the CS method to reduce 2018.8648040.
[12] X. Fan, H. Zhang, C. Leung, Z. Shen, Robust unobtrusive fall detection using
the number of samples and recover the signal using few samples via the infrared array sensors, in: 2017 IEEE International Conference on Multisensor
implementation of the MP algorithm. This design is implemented and Fusion and Integration for Intelligent Systems, MFI, 2017, pp. 194–199, http:
performed on the Zynq-7000 XC7Z020-CLG484-1. The Vivado tools of //dx.doi.org/10.1109/MFI.2017.8170428.
Xilinx are employed to create, synthesize and implement the proposed [13] I.W.W. Wisesa, G. Mahardika, Fall detection algorithm based on accelerometer
and gyroscope sensor data using recurrent neural networks, IOP Conf. Ser. Earth
design on Zedboard. Environ. Sci. 258 (2019) 012035, http://dx.doi.org/10.1088/1755-1315/258/1/
Four IPs were built and simulated during the implementation’s 012035.
workflow: Compress, Recovery, Features and Predict. The implemen- [14] C.-L. Lin, W.-C. Chiu, T.-C. Chu, Y.-H. Ho, F.-H. Chen, C.-C. Hsu, P.-H. Hsieh,
C.-H. Chen, C.-C.K. Lin, P.-S. Sung, P.-T. Chen, Innovative head-mounted system
tation was made according to two approaches; with optimization and
based on inertial sensors and magnetometer for detecting falling movements,
without any optimization. In this last approach, the results confirm Sensors 20 (20) (2020) http://dx.doi.org/10.3390/s20205774, URL https://
that the system is operating with the same performance recognition as www.mdpi.com/1424-8220/20/20/5774.
founded on software and performs better in terms of execution time. [15] E. Boutellaa, O. Kerdjidj, K. Ghanem, Covariance matrix based fall detection
from multiple wearable sensors, J. Biomed. Inform. 94 (2019) 103189, http:
In the second approach, we apply various optimization (Pipeline) to
//dx.doi.org/10.1016/j.jbi.2019.103189, URL https://www.sciencedirect.com/
the three IPs (Compress, Recovery, and Features). This last approach science/article/pii/S1532046419301078.
achieved very interesting performance in terms of latency where the [16] S. Yousuff, S.K. Chaudary, N.P. Meghana, T.S. Ashwin, G. Ram Mohana Reddy,
Compress module is 46x faster; the Recovery block is 5X faster and Zigbee-based wearable device for elderly health monitoring with fall detection,
in: P.K. Sa, S. Bakshi, I.K. Hatzilygeroudis, M.N. Sahoo (Eds.), Recent Findings
the Features block is 3X faster as compared to the no optimization
in Intelligent Computing Techniques, Springer, Singapore, 2018, pp. 69–76.
approach. [17] D. Yacchirema, J.S. de Puga, C. Palau, M. Esteve, Fall detection system for
As future work, we will focus on implementing other classification elderly people using IoT and big data, in: The 9th International Conference
algorithms such as SVM, Decision Trees, etc. We intend to use data on Ambient Systems, Networks and Technologies (ANT 2018) / The 8th
International Conference on Sustainable Energy Information Technology (SEIT-
fusion techniques to reduce computation. Our architecture can be
2018) / Affiliated Workshops, Procedia Comput. Sci. 130 (2018) 603–610,
extended as a monitoring application for a patient to process physio- http://dx.doi.org/10.1016/j.procs.2018.04.110, URL https://www.sciencedirect.
logical signals (oxygen saturation, blood glucose, etc.). com/science/article/pii/S1877050918304721.

6
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

[18] A. Tahir, G. Morison, D.A. Skelton, R.M. Gibson, Hardware/software Co-design of Kerdjidj Oussama obtained his Ph.D. degree at the Uni-
fractal features based fall detection system, Sensors 20 (8) (2020) http://dx.doi. versity of Laghouat, Algeria, in 2019. His research interests
org/10.3390/s20082322, URL https://www.mdpi.com/1424-8220/20/8/2322. involve embedded systems, high-performance computing us-
[19] O. Kerdjidj, N. Ramzan, K. Ghanem, A. Amira, F. Chouireb, Fall detection and ing FPGAs, Hardware and Software implementation, signal
human activity classification using wearable sensors and compressed sensing, J. processing, and Artificial Intelligence applied to Healthcare
Ambient Intell. Humaniz. Comput. 11 (1) (2020) 349–361. applications. He is currently researcher at the Center for
[20] R.M. Gibson, A. Amira, N. Ramzan, P.C. de-la Higuera, Z. Pervez, Matching Development of Advanced Technologies (CDTA).
pursuit-based compressive sensing in a wearable biomedical accelerometer fall
diagnosis device, Biomed. Signal Process. Control 33 (Supplement C) (2017)
96–108, http://dx.doi.org/10.1016/j.bspc.2016.10.016.
[21] H. Djelouat, H. Baali, A. Amira, F. Bensaali, CS-based fall detection for connected
health applications, in: 2017 Fourth International Conference on Advances in
Biomedical Engineering, ICABME, 2017, pp. 1–4.
[22] O. Kerdjidj, A. Amira, K. Ghanem, N. Ramzan, S. Katsigiannis, F. Chouireb, Elhocine Boutellaa received Ph.D. degree in Computer
An FPGA implementation of the matching pursuit algorithm for a compressed Science from Ecole Nationale Supérieure dInformatique,
sensing enabled e-health monitoring platform, Microprocess. Microsyst. 67 Algeria. He has been working as research associate with
(2019) 131–139, http://dx.doi.org/10.1016/j.micpro.2019.03.007, URL http:// Centre de Développement des Technologies Avancées since
www.sciencedirect.com/science/article/pii/S0141933118304630. 2010. Currently he holds the position of postdoctoral re-
[23] H. Djelouat, X. Zhai, M. Al Disi, A. Amira, F. Bensaali, System-on-chip solution searcher at University of Oulu. His research interests include
for patients biometric: A compressive sensing-based approach, IEEE Sens. J. 18 face analysis, biometrics, and healthcare and computer
(23) (2018) 9629–9639, http://dx.doi.org/10.1109/JSEN.2018.2871411. vision.
[24] H. Rabah, A. Amira, A. Ahmad, Design and implementaiton of a fall detection
system using compressive sensing and shimmer technology, in: 2012 24th
International Conference on Microelectronics, ICM, 2012, pp. 1–4, http://dx.doi.
org/10.1109/ICM.2012.6471399.
[25] H. Djelouat, H. Baali, A. Amira, F. Bensaali, CS-based fall detection for connected Abbes Amira (Senior Member, IEEE) received the Ph.D.
health applications, in: 2017 Fourth International Conference on Advances in degree in computer engineering from Queen’s University
Biomedical Engineering, ICABME, 2017, pp. 1–4. Belfast, U.K., in 2001. Since 2001, he has been taking
[26] H. Djelouat, M. Al Disi, I. Boukhenoufa, A. Amira, F. Bensaali, C. Kotronis, many academic and consultancy positions in U.K., Europe,
E. Politi, M. Nikolaidou, G. Dimitrakopoulos, Real-time ECG monitoring using Asia, and Middle East. In U.K., he has previously taken
compressive sensing on a heterogeneous multicore edge-device, Microprocess. academic and leadership positions at Queen’s University
Microsyst. 72 (2020) 102839, http://dx.doi.org/10.1016/j.micpro.2019.06.009, Belfast, Brunel University London, and the University of
URL https://www.sciencedirect.com/science/article/pii/S0141933119300390. Ulster. From 2017 to 2019, he was an Associate Dean for
[27] O. Kerdjidj, N. Ramzan, A. Amira, K. Ghanem, F. Chouireb, Design and research and graduate studies with the College of Engineer-
evaluation of vivado HLS-based compressive sensing for ECG signal analysis, ing, Qatar University, Qatar. He has taken visiting professor
in: 2018 IEEE 16th Intl Conf on Dependable, Autonomic and Secure Computing, positions at the University of Tun Hussein Onn, Malaysia,
16th Intl Conf on Pervasive Intelligence and Computing, 4th Intl Conf on Big and the University of Nancy, Henri Poincare, France. He
Data Intelligence and Computing and Cyber Science and Technology Congress, is currently an Associate Dean for research and innovation
DASC/PiCom/DataCom/CyberSciTech, IEEE, 2018, pp. 457–461. with the Faculty of Computing, Engineering and Media,
[28] J. Li, P. Chow, Y. Peng, T. Jiang, FPGA implementation of an improved OMP for De Montfort University, U.K. During his career to date, he
compressive sensing reconstruction, IEEE Trans. Very Large Scale Integr. (VLSI) has been successful in securing substantial funding from
Syst. 29 (2) (2020) 259–272. government agencies and industry. He has supervised more
[29] W. Saadeh, M.A.B. Altaf, M.S.B. Altaf, A high accuracy and low latency patient- than 25 Ph.D. students and has more than 350 publications
specific wearable fall detection system, in: 2017 IEEE EMBS International in top journals and conferences in the area of embedded
Conference on Biomedical Health Informatics, BHI, 2017, pp. 441–444, http: systems, the IoT, image, and signal processing. He has
//dx.doi.org/10.1109/BHI.2017.7897300. also conducted consultancy services for several government
agencies and companies in the private sector. His research
[30] X. Song, T. Xie, S. Fischer, A memory-access-efficient adaptive implementation
interests include embedded systems, high-performance com-
of KNN on FPGA through HLS, in: 2019 IEEE 37th International Conference
puting, big data and the IoT, connected health, image
on Computer Design, ICCD, 2019, pp. 177–180, http://dx.doi.org/10.1109/
and vision systems, biometric and security., Prof. Amira
ICCD46524.2019.00030.
has participated as a member and a Guest Editor of the
[31] Z.-H. Li, J.-F. Jin, X.-G. Zhou, Z.-H. Feng, K-nearest neighbor algorithm imple-
editorial board in many international journals including
mentation on FPGA using high level synthesis, in: 2016 13th IEEE International
recent special issues in the IEEE IoT Journal and Pattern
Conference on Solid-State and Integrated Circuit Technology, ICSICT, 2016, pp.
Recognition (Elsevier). He is a Fellow of IET and of the
600–602, http://dx.doi.org/10.1109/ICSICT.2016.7998989.
Higher Education Academy, and a Senior Member of ACM.
[32] L.H. Crockett, R.A. Elliot, M.A. Enderwitz, R.W. Stewart, The Zynq Book:
He received many international awards, including the 2008
Embedded Processing with the Arm Cortex-A9 on the Xilinx Zynq-7000 All
VARIAN prize offered by the Swiss Society of Radiobiology
Programmable Soc, Strathclyde Academic Media, UK, 2014.
and Medical Physics, the CAST Award, the IET Premium
[33] A. Burns, E.P. Doheny, B.R. Greene, T. Foran, D. Leahy, K. O’Donovan, M.J.
Award, in 2017, the DELL-EM Envision the Future, in 2018,
McGrath, An extensible platform for physiological signal capture, in: Engineering
and many best paper and recognition awards in the IEEE
in Medicine and Biology Society, EMBC, 2010 Annual International Confer-
international conferences and events.
ence of the IEEE, 2010, pp. 3759–3762, http://dx.doi.org/10.1109/IEMBS.2010.
5627535.
[34] D.L. Donoho, Compressed sensing, IEEE Trans. Inform. Theory 52 (4) (2006) Khalida Ghanem received the B.Sc. in electronics from École Nationale Polytechnique
1289–1306, http://dx.doi.org/10.1109/TIT.2006.871582. (ENP) of Algiers and the M.Sc. and Ph.D. degrees in telecommunications from Institut
[35] E.J. Candes, M.B. Wakin, An introduction to compressive sampling, IEEE Sig- National de la Recherche Scientifique (INRS), University of Quebec, Canada in 2008
nal Process. Mag. 25 (2) (2008) 21–30, http://dx.doi.org/10.1109/MSP.2007. and the habilitation (HDR) in 2012 from ENP. From 1999 to 2000, she was with
914731. the Department of Mathematics, Electrical and Computer Engineering, University of
[36] Y. Arjoune, N. Kaabouch, H. El Ghazi, A. Tamtaoui, Compressive sensing: Quebec at Rimouski, in Canada, where she worked as a Research Assistant on practical
Performance comparison of sparse recovery algorithms, in: 2017 IEEE 7th Annual smart antenna systems. From 2008 to 2010, she joined the School of Electronic,
Computing and Communication Workshop and Conference, CCWC, 2017, pp. Electrical and Computer Engineering, University of Birmingham, in U.K., as a Research
1–7, http://dx.doi.org/10.1109/CCWC.2017.7868430. Associate, and the University of Southampton as a Research Visitor, and then the
[37] O. Kerdjidj, K. Ghanem, A. Amira, F. Harizi, F. Chouireb, Real ECG signal University of Sheffield, as a Research Fellow. From 2010 to 2011, she was an
acquisition with shimmer platform and using of compressed sensing techniques Assistant Professor with Prince Mohammad Bin Fahd University, KSA. Since 2011,
in the offline signal reconstruction, in: 2016 IEEE International Symposium on she has been with the Center for Development of Advanced Technologies (CDTA),
Antennas and Propagation, APSURSI, 2016, pp. 1179–1180, http://dx.doi.org/ in Algeria, where she is a Research Director. In 2012 she founded the wireless
10.1109/APS.2016.7696297. communication group and has acted as its leader since then. From 2012–2020, she
was a Principal Investigator of multiple projects funded by the General Direction of
Scientific Research and Technological Development (DGRSDT), conducting research in
ultrawideband communications, resource allocation in cognitive radios, 5G-and-beyond

7
O. Kerdjidj et al. Microprocessors and Microsystems 91 (2022) 104514

wireless systems, and body area networks. She was Chair of IEEE women in engineering Fatima Chouireb received her Electrical Engineering
(WIE) group in Montreal and its vice-Chair in KSA. She is an Adjunct Professor diploma in 1992, and her Magister and Ph.D. degrees from
at the department of Electrical Engineering at the university of Quebec at Abitibi- Saad Dahlab University of Blida, Algeria in 1996 and 2007
Temiscamingue in Canada. She is also the recipient of excellence fellowships and respectively. In 1997, she joined the Electrical Engineering
awards and acts as an expert for research and governmental institutions. Her research Department of Laghouat University, Algeria, as Assistant
interests lie in the field of wireless communications, channel characterization and Lecturer. Since December 2007, Dr. Chouireb is an Assistant
modeling and sensor-based systems for e-health and sport applications and encompass Professor at the same Department. She is also a Team Leader
body and In-vivo area networks, design of multiple antenna (MIMO) systems, and of ‘‘Signal, Image and Speech’’ research group of LTSS
cognitive and cooperative communication techniques. Laboratory, Laghouat University, Algeria. She is currently a
Full Professor at the same Department. Her current research
interests include: Signal, Image and Speech processing,
computer vision, Localization/Mapping/SLAM and mobile
robotics problems