An Improved Deep-learning Network for Abnormal Action Detection
Jiawei Zhao, Jia Qiao, Lijun Feng, Yong Zhang · 2022
Detecting actions in untrimmed videos is a challenging task. In order to improve the accuracy of abnormal action detection, an improved network combining Boundary-Matching Network (BMN) and Structured Segment Network (SSN) is proposed in this paper to achieve the temporal detection of abnormal actions from public places. BMN is used to generate the temporal proposals of abnormal actions in the video, and then SSN acts on the proposals generated by BMN to classify them into specific categories. By modifying the feature dimensions to adapt to the length of the video, and transforming the output generated by the BMN network into a proposal, the BMN and SSN networks are well combined. The experimental results prove that the proposed method achieves good results on the abnormal action dataset collected from public places.