Video Anomaly Detection with Spatio-Temporal Inspired Deep Neural Networks (DNN)

Intelligent surveillance systems must be able to detect anomalies promptly to prevent malicious activity. It is common for deep learning methods to be used in video anomaly detection to focus on analyzing video streams from just one camera with a single scenario. Using large-scale training data with high complexity is necessary for these deep learning methods. This paper uses a spatiotemporal-inspired Deep Neural Network (DNN) to detect video anomalies. Rather than expensive optical flow calculations, a Deep Neural Network (DNN) is used for motion information in the proposed approach to achieve high recognition accuracy at a low computational cost. Experimental results on publicly available datasets demonstrate that the proposed model provides better input frame generation performance and is more accurate than existing approaches.