Abnormal behavior detection in video using BiFPN-EfficientNet based on cosine distance optimization
Haili X. Zhao, Junfang Song, Xiaoyu Xu, Tengjiao Wang, Yuanyuan Pu, Wenzhe Wu · 2023
With the explosive growth of the number of videos on surveillance and the Internet. How to intelligently detect the large amount of video data has become a core task in the current video security surveillance system and network security, which is of great significance to urban security control and network environment governance. This paper introduce the idea of metric learning and design a video abnormal behavior detection network EBML applied to small sample size to address the problem that specific abnormal behavior cannot be detected and abnormal behavior samples are scarce in most methods. Firstly, EfficientNet is used as the backbone network for feature extraction, then, bringing in an improved BiFPN to mitigate the feature loss caused by too many networks and information loss, and enhances the ability to fuse low-level semantic information. Finally, the scalable Cosine distance metric is introduced into the Softmax of BiFPN-EfficientNet, which makes the gap between similar features shrink continuously and the gap between dissimilar features increase continuously, thus improving the accuracy of video anomalous behavior detection. By testing on the VCAD dataset, the accuracy reaches 96.3% and the F1 value is 0.94, which has an increase of 7.8% and 0.09 respectively, compared with the benchmark model EfficientNet. Experiments show that the algorithm can better balance speed and accuracy in video anomaly detection and can meet the demand of video anomalous behavior detection under small samples.