Abnormal Behavior Recognition Based on Hybrid Depth Model with SE Block
Yuanyao Lu, Dongjie Li, Jingxuan Wang · 2022 4th International Conference on Advances in Computer Technology, Information Science and Communications (CTISC) · 2022
The detection and recognition of human abnormal behavior in surveillance video based on deep learning has become a research hotspot. Due to the complex background in the video and the variety of targets, it is difficult to accurately identify abnormal behavior in surveillance video. The current abnormal behavior recognition models have some shortcomings such as complex structure, big calculation cost and low recognition rate. Given the aforementioned obstacles, this paper proposes an ISE-R3D-LSTM network based on R3D model for abnormal behavior detection in surveillance video, which integrates the improved SE block and LSTM network. The contribution of the proposed algorithm consists of three parts: First, considering the global information extraction and feature calibration, the improved SE block with higher accuracy is obtained by substituting the FC layer of the conventional SE block with the GAP; Second, a residual module is added to the 3D CNN so as to avoid gradient dispersion, and the GAP is utilized to reduce the number of the network parameters; Finally, to effectively learn feature information, LSTM is introduced to carry out the timing modeling for high-level features, yielding our ISE-R3D-LSTM network. The RGB-only video data from UCF Crime dataset is employed for the experimental evaluation. The experimental results present that the recognition rate of our ISE-R3D-LSTM network reaches 92.30%. Comparative results present that our model is competitive in terms of computation cost and recognition accuracy. (Abstract)