Development of Anomalous Video Detection System using Hybrid-Features Analysis of Actions and Scene-Backgrounds Information

Abstract

Detecting an anomalous event in video clips captured from a surveillance camera is an important task, especially for security system purposes. However, as the probability of such occurrence is very low, automatic detection of the anomalous event is then necessary to replace the human labor-intensive works. Researchers have developed various methods to solve this problem, however, most of the proposed methods limit their definitions of anomalous events that might be perceived as having a different meaning when it occurs in other scene backgrounds. We have developed an automatic anomalous video detection system by extracting the individual action from the video clips, followed by extracting also various scene-background characteristics related with the respective action, and represented as a Video Graph to be classified as an anomaly through a Graph Convolutional Networks. We also constructed a new database as the available databases could not be used in this experiment. Results of experiments show that the various anomalous actions videos have been successfully detected with higher recognition capability compared with that of the conventional methods.