MULTI-CAMERA BASED INTELLIGENT VIDEO SURVEILLANCE SYSTEM USING SSD_MOBILENET_V3 | ||
| Journal of Al-Azhar University Engineering Sector | ||
| Volume 20, Issue 76, July 2025, Pages 974-996 PDF (2 M) | ||
| Document Type: Original Article | ||
| DOI: 10.21608/auej.2025.351357.1752 | ||
| Authors | ||
| M Shaban* 1; Marwa Elpeltagy1; Ahmed Y Khedr1; A Al-Marakeby2 | ||
| 1Systems and Computer Engineering Department, Faculty of Engineering, Al-Azhar University, Cairo, Egypt | ||
| 2Systems and Computers Engineering Dept. , Faculty of Engineering , Al-Azhar University, Cairo, Egypt. | ||
| Abstract | ||
| Contemporary intelligent surveillance systems have shifted from passive monitoring to active threat detection through advanced behavioral analytics. This work introduces a flexible, high-speed, real-time anomaly detection framework that dynamically evaluates critical security threat indicators, including spatiotemporal event patterns, object-class kinematics, trajectory semantics, and scene-context deviations. The system detects both abnormal human activities (e.g., unauthorized zone intrusions, sudden locomotor anomalies) and anomalous vehicular events (e.g., contraflow violations, perimeter breaches) through integrated spatiotemporal analysis. A graphical user interface (GUI) allows users to define abnormal scenarios. Users can customize abnormal factors, including event time, object type, and direction of abnormal movement. They also can draw areas of interest. The deep neural network Single Shot Multi-Box Detector (SSD_MobileNet_v3) is employed to detect objects within video frames. Subsequently, a Kernelized Correlation Filter (KCF) tracker is used to monitor these objects and identify abnormal motion direction. An innovation of this system is its ability to classify human motion types by establishing a relationship between the actual dimensions of individuals and the observed distances in the video. Furthermore, a dataset for abnormal behavior detection has been created. The efficiency of the system is assessed on the authentically distorted surveillance video dataset and the obtained dataset. The proposed method achieves a recall rate of 95.83% for truck detection and 92.42% for walking intrusion detection. An F1_score of 87% for motion classification is reported. An average processing speed of 16 frames per second (fps) has been scored. The research results show that the suggested strategy outperforms existing state-of-the-art procedures | ||
| Keywords | ||
| Multi-camera Video Surveillance; Direction Detection; SSD_MobileNet_v3; Human Motion Classification | ||
|
Statistics Article View: 116 PDF Download: 76 |
||
