Deep Learning-Based Alert Generation for

Detecting Suspicious Activity

Prof. Amit Bhosale
Prof. Sagar Patole Prof. Ms.Sayali Alate
Computer Engineering Department
Computer Engineering Department Computer Engineering Department
PES’s Collegeof Engineering
PES’s Collegeof Engineering PES’s Collegeof Engineering
Phaltan, India
Phaltan, India Phaltan, India
[email protected]
[email protected] [email protected]
 Abstract— In the today’s world, video surveillance is of indicative of suspicious behavior. By leveraging deep
utmost importance. The integration of Deep Learning has learning techniques, the proposed approach achieves high
significantly advanced surveillance technologies. Various accuracy in detecting various types of suspicious activities,
systems have been developed to differentiate between different
thereby enhancing security surveillance systems' efficacy.
suspicious behaviors observed in live footage. Human behavior,
being highly unpredictable, presents a challenge in
Tejashri Subhash Bora [3] proposed Network (CNN)
determining its nature as suspicious or normal. To tackle this models for video surveillance applications. The primary
challenge, Deep Learning is employed to detect suspicious or goal is to identify and alert shop owners about suspicious
normal activity within academic environments. When a activities observed in surveillance footage. The authors
suspicious activity is predicted, an alert message is sent to the emphasize the significance of automated surveillance
relevant authority. Monitoring typically involves analyzing systems in enhancing security measures and reducing
consecutive frames extracted from the video feed. manual efforts required for monitoring. By leveraging CNN
models, the proposed system can effectively analyze video
Keywords—CNN Algorithm, Suspicious activity, Video
data and detect various suspicious behaviors accurately. The
capture.
study underscores the potential of deep learning techniques
in improving the efficiency and effectiveness of video
I. INTRODUCTION surveillance systems for ensuring public safety.
Detecting human behaviour in real-world environments
has numerous applications, spanning from intelligent video
surveillance to analyzing shopping behaviors. Video IV. PROPOSED METHODOLOGY
surveillance is widely applicable across indoor and outdoor The proposed system aims to utilize video footage obtained
settings, serving as a crucial component of security from cameras for monitoring activities in places such as
measures. Security cameras have become ubiquitous for banks or bus stations. It will promptly send alert messages to
safety and security purposes, integrating seamlessly into nearby police stations whenever any suspicious events are
daily life. E-surveillance stands as a prominent agenda within
detected. In this project, we focus on determining whether a
the Digital India development program initiated by the
person's behavior is suspicious or not. Nowadays, CCTV
Indian government, with video surveillance remaining a
pivotal aspect of this initiative. The proposed model will be cameras are ubiquitously installed, capturing videos stored
transformed into a desktop application capable of analyzing on centralized servers. However, manually scanning these
CCTV footage as input and issuing alerts to administrators videos to detect suspicious activity demands significant
upon detecting suspicious poses. Identifying suspicious human effort and time. To address this challenge, we
human activity entails recognizing specific human body parts propose automating this process using Deep Learning
and potentially tracking their movements. Algorithms. Specifically, convolutional neural networks will
be employed to identify potential suspicious activities.
II. PROBLEM DEFINITION A. Algorithm
Detecting suspicious human activity from surveillance video CNN: Convolutional Neural Networks (CNNs) exploit
is a dynamic research field within image processing and spatial correlations inherent in input data. In CNNs, each
computer vision. Video surveillance enables the monitoring layer of the neural network connects certain input neurons,
of human activities in both sensitive and public locations. forming what is known as a local receptive field. This field
This project aims to utilize neural networks to identify concentrates on hidden neurons, enabling efficient
suspicious human activity from video datasets. processing of complex visual information. CNNs represent
specific network architecture within deep learning
III. LITERATURE REVIEW algorithms, primarily tailored for tasks like image
Om M. Rajpurkar [1] focuses on utilizing transfer learning recognition that entail the analysis of pixel data. While
techniques to detect suspicious activities in surveillance various neural network architectures exist in the realm of
footage. The primary objective is to alert shop owners deep learning, CNNs stand out as the preferred choice for
promptly upon detecting such activities. The authors object identification and recognition due to their capability
emphasize the significance of real-time surveillance for to effectively capture and process spatial relationships
enhancing security measures. By employing transfer within visual data.
learning, the model can leverage pre-trained neural networks CNN Algorithm Steps: -
to adapt and learn from new data efficiently. This approach
not only improves detection accuracy but also reduces
 Convolution layer.
computational resources required for training. The study  ReLU layer.
underscores the importance of timely alerts in preventing  Pooling layer.
potential security threats, thereby enhancing safety in  Fully connected layer.
commercial establishments. Amruta C. V [2] presents a deep
learning-based approach for detecting suspicious activities Convolution layer :The convolutional layer serves as the
in surveillance videos. The primary focus is on fundamental component of a CNN. Comprising a collection
distinguishing between suspicious and normal activities, of filters, also known as kernels, this layer's parameters are
with the system promptly alerting the relevant authorities adjustable and refined during the training process. Typically
upon detection of suspicious behavior. The authors highlight smaller in size compared to the input image, each filter
the effectiveness of deep learning algorithms in handling undergoes convolution with the image, generating an
complex visual data and capturing intricate patterns activation map as a result.
ReLU layer: As a result, employing ReLU helps mitigate such footage from students may pose challenges, support
the exponential surge in computational demands for from civic authorities could facilitate access to footage
operating the neural network. When the CNN expands, the documenting criminal incidents that have occurred in the
computational expense of incorporating additional ReLUs recent past. This collaboration would significantly contribute
escalates linearly. to refining and strengthening the efficacy of the model.
Pooling layer: Pooling layers serve to decrease the
dimensions of feature maps, thereby diminishing the number
VI. CONCLUSION
of parameters to be learned and the computational workload
within the network. These layers effectively summarize the This system enables the detection of suspicious human
features within specific regions of the feature map produced activities and can be expanded to analyze and interpret
by a convolutional layer. people's behavior in various locations. The outcome of this
system is the ability to identify whether any suspicious
Fully connected layer: A fully connected layer denotes a
activity is occurring. In today's society, the significance of
neural network where each neuron applies a linear
CCTV footage is widely recognized, albeit often utilized for
transformation to the input vector using a weight matrix.
investigative purposes post-incident. In contrast, the
Consequently, all potential connections from layer to layer
proposed model offers the advantage of proactively
exist, ensuring that every input within the input vector
preventing crimes. It operates by continuously monitoring
impacts every output within the output vector.
and analyzing real-time CCTV footage. Upon analysis, if any
potential untoward incident is detected, the system promptly
alerts the relevant authorities to take necessary action, thus
B. Image Processing
averting the potential crime. Hence, the proposed system
Image preprocessing involves the procedures undertaken to represents a machine-based approach to identify criminal
format images before they are utilized for model training. activities in surveillance videos, addressing the growing need
Such preparation is essential to ready picture data for model for enhanced security measures amidst the rising incidence
input. of daily crimes. The system's outcome lies in its ability to
C. Feature Extraction identify anomalous actions in real-time.
Feature extraction involves the process of isolating only REFERENCES
the essential features from an image and producing the
output by eliminating any noise or unnecessary
interruptions. [1] Om M. Rajpurkar, Siddesh S. Kamble, Jayram P.
Nandagiri, and Anant V. Nimkar, “ALERT
D. Image Classification GENERATION ON DETECTION OF SUSPICIOUS
ACTIVITY USING TRANSFER LEARNING”
Image classification is a fundamental task in computer vision
aiming to interpret an entire image in its entirety. The 2020@IEEE Department of Computer Engineering
objective is to categorize the image as either suspicious or Sardar Patel Institute of Technology Mumbai, India
not. IEEE – 49239
[2] Amrutha C.V, C. Jyotsna, Amudha J. “Deep
E. Dataset Collection Learning Approach for Suspicious Activity Detection
The dataset consists of video files in the MP4 format, from Surveillance Video”2020@IEEE Dept. of
categorized into suspicious and normal forms. The videos Computer Science & Engineering, Amrita School of
depict various scenarios such as fighting, fire incidents, Engineering,Bengaluru, Amrita Vishwa Vidyapeetham,
attacks, and more (ICIMIA 2020) IEEE Xplore Part Number: CFP20K58-
ART; ISBN: 978-1-7281-4167-1.
F. Training and Testing Model
Training and testing are techniques employed to evaluate the [3] Tejashri Subhash Bora, Monika Dhananjay Rokade.
accuracy of a model, collectively known as Train/Test. The “HUMAN SUSPICIOUS ACTIVITY DETECTION
training data comprises a subset of the original dataset SYSTEM USING CNN MODEL FOR VIDEO
utilized for training the model, while the testing data is SURVEILLANCE” 2021@ijariie
employed to assess the model's accuracy. PG Student, Department of Computer Engg., SPCOE,
Maharashtra, India Assistant Prof., Department of
Computer Engg., SPCOE, Maharashtra, India Vol-7
V. FUTURE SCOPE Issue-3 2021 IJARIIE- ISSN(O)-23954396.
Identifying suspicious activity in public spaces like bus [4] Nipunjita Bordoloi, Anjan Kumar Talukdar,
stations and banks is the primary focus of our application, Kandarpa Kumar Sarma, “Suspicious Activity Detection
designed for user-friendly access. Currently, our model is from Videos using YOLOv3” Dept. of Electronics and
tailored to detect only five specific suspicious activities. Communication Engineering Gauhati University,
However, there is room for enhancement by expanding its Guwahati, Assam, India 2020@IEEE. 17th India Council
scope to include a wider range of suspicious behaviors. This International
can be achieved by augmenting the dataset with additional [5] Rachana Gugale, Abhiruchi Shendkar, Arisha
images, particularly those extracted from CCTV footage Chamadia, Swati Patra, Deepali Ahir. “Human
capturing instances of suspicious activity. While obtaining Suspicious Activity Detection using Deep Learning”.
Student, Department of Computer Engineering, M. E. Video using Discriminative Deep Belief Netwok”,
S. College of Engineering, Pune, India. Assistant International Journal of Control Theory and
Professor, Department of Computer Engineering, M. Applications Volume 10, Number 29
E. S. College of Engineering. Volume: 07 Issue: 06 -2017.
June 2020@irj.
[6] J. R. Medel and A. Savakis, “Anomaly detection U.M.Kamthe,C.G.Patil “Suspicious Activity
in video using predictive convolutional long short-term Recognition in Video Surveillance System”, Fourth
memory networks,” arXiv preprint arXiv:1612.00390, International Conference on Computing Communication
2016. Control and Automation (ICCUBEA), 2018.
[7] Monika D.Rokade, Dr. Yogesh Kumar Sharma, [11] H. Zhong, J. Shi and M. Visontai, “Detecting
"Deep and Machine Learning Approaches for unusual activity in video”, In Proc. 2004 IEEE Comput
Anomaly-Based Intrusion Detection of Imbalanced Vis. Pattern Recog., Vol. 2,
Network Traffic", IOSR Journal of Engineering, 2019.
pp. 819-826, 2014.
[8] Yuke Li “A Deep Spatiotemporal Perspective for
Understanding Crowd Behavior”, IEEE Transactions
on multimedia, Vol. 20, NO. 12, December 2018. [13]B.Benjdira,T.Khursheed,A.Koubaa,A.Amma
randK.Ouni,”Car Detection using Unmanned Aerial
[9] G. Sreenu and M. A. Saleem Durai “Intelligent Vehicles: Comparison between Faster R- CNN and
video surveillance: a review through deep learning YOLOv3,” 2019 1st International Conference on
techniques for crowd analysis” , Journal Big Unmanned Vehicle Systems- Oman (UVS), Muscat,
Data ,2019. Oman, 2019, pp. 1-6.
[10] Elizabeth Scaria, Aby Abahai T and Elizabeth
Isaac, “Suspicious Activity Detection in Surveillance