Smart Surveillance System for Monitoring and Detecting Abnormal Human Behaviour

Abstract

The increasing demand for real-time video surveillance systems has underscored the need for automated methods to detect abnormal human behaviors, particularly to enhance the safety of individuals, including those living alone. This study presents a deep learning-based framework for identifying anomalous human activities in video data, enabling timely interventions.By leveraging advanced deep CNNs, the proposed framework extracts discriminative features directly from raw video frames, eliminating dependence on manual design features. To mitigate challenges associated with limited labeled training data, transfer learning from large-scale action recognition datasets is employed. The model is trained and rigorously evaluated on a dataset encompassing four critical categories of abnormal behavior: fighting, explosions, robbery, and assault. The project results to show the framework\'s effectiveness in accurately detecting these activities, contributing to enhanced surveillance and public safety.

Introduction

This research focuses on predictive modeling of abnormal human behaviors in surveillance environments, leveraging AI techniques like Convolutional Neural Networks (CNN) to improve over traditional manual monitoring. It emphasizes real-time detection, pattern recognition, and proactive security responses through analysis of behavioral and environmental data.

Key Objectives and Methodology

• Build a model to detect and flag atypical human behavior using CNNs, improving security through real-time alerts.

• Conducted a systematic literature review of recent studies from major databases.

• Focused on high-accuracy models and promising methods like LSTM, YOLO, ResNet, and transfer learning.

• Employed datasets such as UCF-Crime, CAVIAR, PETS, UCSD, V6, Crowd-11, and others for benchmarking.

Highlights from Literature Survey

1. CNN + NLP: Used for behavior analysis on social networks based on profile and URL data (e.g., Facebook, PhishTank).

2. Video Surveillance with 3D CNNs: Identified abnormal activities like loitering, falling, panic, violence, and sexual abuse more effectively than 2D CNNs.

3. Transfer Learning with Drones: Applied modified ResNet-18 models using drone footage for crowd monitoring, achieving over 90% accuracy.

4. 1D CNN for Time-Series: Detected motion anomalies through a weakly supervised learning method.

5. Image-Based Recognition: IHAR system used PCA and various ML models to classify actions (e.g., walking, waving).

6. LSTM + PSO: Addressed real-time Human Activity Recognition (HAR) from video for healthcare and surveillance.

7. YOLO-Based Detection: Applied to patient monitoring, theft detection in schools, and crowd behavior analysis; accuracy up to 99.5% in some cases.

8. Darknet & CNN Models: Used for detecting threats such as weapons, child abuse, and suspicious activities with accuracies as high as 97.39%.

9. Graph Neural Networks: Used with OpenPose to extract skeleton-based features for behavior classification.

10. Transformer-Based Models: For crowd behavior categorization and multi-modal spatial-temporal analysis.

Design and Implementation

Design:

• Utilized pre-trained CNN models through transfer learning.

• Added custom layers for anomaly detection, extracting spatial-temporal features of human behavior.

• Used labeled datasets divided into training, validation, and testing subsets to fine-tune models and prevent overfitting.

Implementation:

• Applied the UCF Crime Dataset for training, covering behavior classes like Assault, Robbery, Explosion, Fighting, Normal.

• Preprocessed data to standardize frames, filter noise, and extract relevant video images (every 10th frame).

• Achieved high performance in detecting abnormal behaviors in real-time settings.

Conclusion

This project presents a robust framework for abnormal behavior detection aimed at enhancing the safety of individuals living alone. By leveraging advanced deep learning architectures, particularly DenseNet201, and employing transfer learning, the model effectively identifies critical abnormal behaviors like fighting, explosions, robbery, and assault in real-time video surveillance. This approach demonstrates high accuracy and computational efficiency, supporting timely interventions and strengthening autonomous monitoring capabilities.Future enhancements could further optimize the system’s effectiveness and broaden its applicability. Refining the current deep learning models and incorporating spatial and temporal analysis would enhance the system’s accuracsy in recognizing complex behavioral patterns over time. The integration of advanced techniques, such as video-based visual transformers, could provide greater insightsintoabnormalbehaviorcontextsandimproveanomalymanagement.Expandingthescopeto include real-time detection for sensitive scenarios, such as patient monitoring and naturalistic video analysis, could also amplify the system’s value in diverse applications. These advancements would not only improve the system’s robustness and adaptability but also establish it as a versatile tool for ensuring the safety of vulnerable individuals in both public and private settings

References

[1] M. Li, H. Zhang, F. Wang. Deep learning models for real-time surveillance and anomaly detection. Journal of Machine Learning Research, 2023;24(1):225-234. doi:10.1162/jmlr.2023.24.1.0106. [2] S. Patel, J. Wen, et al. Spatial and temporal dynamics in human activity recognition with CNNs. Pattern Recognition Letters, 2021;156:45-53. doi:10.1016/j.patrec.2021.06.004. [3] K. Tan, R. Xu, et al. Video-based visual transformers for advanced surveillance applications. IEEE Transactions on Image Processing, 2023;32:1245-1257. doi:10.1109/TIP.2023.3240987. [4] R. Singh, M. Shah, S. Yadav. Transfer learning for low-data anomaly detection in video surveillance. Journal of Computer Vision and Applications, 2022;61:84-96. doi:10.1007/s11127- 022-0987-5. [5] A. Nair, P. Sinha. Temporal analysis in CNNs for real-time abnormal behavior detection. ACM Computing Surveys, 2021;54(8):90-99. doi:10.1145/3460782. [6] S. Roy, V. Khare. Patient safety monitoring through abnormal activity recognition. Computer Methods and Programs in Biomedicine, 2022;217:106787. doi:10.1016/j.cmpb.2022.106787. [7] J. Smith, M. Luo, et al. A comparative analysis of deep learning models for anomaly detection. IEEE Transactions on Cybernetics, 2022;52(4):564-573. doi:10.1109/TCYB.2022.3150456. [8] B. Ahmed, T. Wilson. Abnormal event recognition in surveillance videos using CNNs. Future Generation Computer Systems, 2023;141:234-244. doi:10.1016/j.future.2022.04.001. [9] L.Kaur,D.Agarwal,etal.Real-timesurveillancewithCNNsforanomalymanagement.Journal of Big Data, 2022;9:56-67. doi:10.1186/s40537-022-00568-9. [10] P.Zhao,A.Jha,etal.Videoanalyticsinreal-timeabnormalbehaviordetectionsystems.Pattern Recognition, 2023;130:1245-1260. doi:10.1016/j.patcog.2023.108876. [11] N. Hassan, F. Tan, et al. Multi-class anomaly detection in surveillance videos. IEEETransactions on Neural Networks and Learning Systems, 2023;15:76-87. doi:10.1109/TNNLS.2023.3150984.

Copyright

Copyright © 2025 J. Sarala Bai, J. Keerthana , K. Jyothi , Y. L. Malathi Latha . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.