Suspicious Activity Detection Using Transfer Learning Based ResNet Tracking from Surveillance Videos

Tracking objects in video surveillance is a challenging task for public security. The advent of the semantic approach has greatly raised the growth of anomaly detection. However, the current anomaly detection methods typically experience problems like inadequate use of movement patterns and inconsistency on various datasets. This research proposed a system to enhance the efficiency of anomaly detection in video surveillance. The proposed system consists of two parts that involves object tracking and suspicious activity detection. The overall framework detects and tracks the abnormal objects in video surveillance. The transfer learning-based ResNet tracking has been used for object tracking. Distance Metric Learning (DML) method has been used for detecting suspicious activities in video surveillance. The results are estimated to analyze the efficiency of the proposed method. The proposed network classifier is compared with the existing ResNet and VGG-16 network. The proposed method provides 99% accuracy that had high performance compared to other existing methods.