Real-Time Surveillance System Using Key Point Detection and Graph Convolutional Networks for Suspicious Activity Recognition

Real-time video surveillance is a critical tool for public safety, enabling crime prevention, crowd management, and timely emergency response. However, traditional surveillance systems face challenges in dynamic environments, including occlusions, low-light conditions, and high labor intensity, which limit their accuracy and reliability. This study proposes an advanced real-time surveillance system integrating OpenPose for key point detection and Graph Convolutional Networks (GCNs) for spatio-temporal activity recognition. The system is based on a modular architecture consisting of preprocessing of video, extraction of key points, and real time activity classification, which is capable of effectively identifying suspicious behaviors. On the COCO Keypoints, MPII Human Pose, and a custom surveillance dataset, the system achieved precision of 91.0%, recall of 88.5% and an F1 score of 89.7 and the latency of 0.48 seconds, therefore making it a suitable candidate for real time applications. Performance analyis under occlusions, low light conditions and dynamic environments showed robust drops of ~12%, ~7% and ~5% respectively demonstrating that robustness needs to be improved. We compared our results with baseline methods (OpenPose + LSTM, CNN + LSTM) and achieved significant improvements in all principal key metrics such as mean average precision (mAP) of 74.5%. These results validate the proposed system as a state-of-the-art real time surveillance system capable of dealing with different operational constraints and at the same time provide timely and accurate detection of suspicious activities. As a result of the system’s modular design, scalability and adaptability, it is suitable for deploying in more complex public safety scenarios.