Human suspicious behavior detection system using deep learning with a closed-circuit television camera

Automated interpretation of human activities in public surveillance is a pressing need. This study introduces a lightweight behavior detection system combining OpenPose (skeletal keypoints) with a Temporal Convolutional Network (TCN) for sequence analysis. The pipeline is implemented on an NVIDIA Jetson Orin Nano edge device, enabling real-time, privacy-preserving processing of anonymized skeletal data. Experiments on the MPII Human Pose dataset show the TCN model surpasses an LSTM baseline, achieving 91.83% overall accuracy and an 8% improvement on high-velocity actions (e.g., running, punching). A six-class confusion matrix showed a good balance, though walking was the most ambiguous category. Field tests in public spaces, using the model as an edge microservice, validated its real-time performance. The system delivers prompt alerts for safety-critical incidents, such as falls or violence. It demonstrates how edge AI can enhance public safety by reducing delays, minimizing data consumption, and diminishing false alarms. The findings indicate TCNs are proficient in comprehending long-term temporal patterns for behavior detection.