Anomaly Detection Using Feature Encoding and Trajectory Association on Edge Devices for Residential Video Surveillance

In recent years, intelligent surveillance becomes one of the most important community services for residential societies. Intelligent surveillance systems detect anomalies in surveillance scenes, which meet the growing demands of security. At residential societies, installation of high-capacity computational equipment for surveillance purpose is found to be very expensive and unfeasible. In this paper, a system is proposed consisting of CPU-only edge devices to detect anomalies. A modular framework is presented to record object-level inferences and tracking for detecting any freak in surveillance scene. Feature encoding and trajectory association guided by two complementary metrices are employed. These are used to deal with partial occlusions, posture deformations, and complicated scenarios. Proposed anomaly detection framework is tailored to run on CPU-only edge devices to achieve desirable Frames Per Second (FPS) parameter that suits for residential surveillance. Experimental results show that the proposed system is viable and produces satisfactory results in real-world situations when compare to other state of the art methods.