Video Anomaly Detection based on Multi-scale Time Network Model
Minyue Fu, Jianxing Zhu, Zhiyong You · 2025
The ability of traditional regression and shallow learning methods to extract data features is inadequate, which leads to the poor effect of anomaly detection. When dealing with long sequence data, deep learning faces great difficulties, especially the phenomenon of gradient disappearance. This means the deep learning model can only make exception judgments when processing such data. Given the above problems, the multi-scale time learning network min (multiple instance learning) realizes the detection of abnormal data in time series. MIN can be used to capture information for long and short periods of time in a video. This method can well enhance the recognition of positive examples, that is the rare abnormal fragments in the abnormal video. Based on this, the multiscale temporal feature learning (MTN) proposed in this paper covers the multi-resolution local time correlation and the global time correlation between video clips. By estimating error sequences, this method detects abnormal data, thereby enhancing algorithm accuracy. Experimental results demonstrate the effectiveness of this method in detecting abnormal data. The algorithm achieves a 0.5% improvement on the UCSD Ped2 dataset, and its effectiveness is further validated on the CUHK Avenue dataset.