Real-Time Anomaly Detection and Feature Analysis Based on Time Series for Surveillance Video

The intelligent surveillance system urgently needs the real-time machine recognition of abnormal events to solve the extremely uneven human supervision resource and digital cameras. Besides, the number of anomaly types that real-time machine monitoring could recognize has not met the need. This paper presents a fast and robust methodology for real-time anomaly detection under different scenarios. We created the Video-Energy-Vector(VEV) to significantly reduce the dimension of feature maps while maintaining the spatial-temporal information. We applied the proposed method on different computer vision features to evaluate the effectiveness of common features to different types of abnormal events based on SVM. Also, we adopted the voting model among different features, which significantly increased the performance. Further More, the small video size to be trained guaranteed the real-time efficiency. The result of the modified UCF-Crime Dataset has proved that our approach has achieved robust results and had the generalization ability on new anomaly types.