基于多尺度量化特征的视频异常行为检测算法

打开文本图片集

中图分类号：TP391.4 文献标志码：A DOI：10.13705/j. issn.1671-6841.2024029

文章编号：1671-6841（2025）05-0039-07

Abstract：Abnormal behavior detection in video had significant application value in the field of surveillance and security. To address the issue of abnormal information generalization caused by skip connections between the encoder and decoder in autoencoder models in generating video frames，an algorithm for video abnormal behavior detection based on multi-scale quantized features was proposed. Firstly，the encoder was trained to learn normal frames and performed vector quantization in a hierarchical manner， while the decoder generated video frames based on the quantized features，avoiding direct information transmission between the encoder and decoder，to significantly reduce the impact of generalization，and to improve the quality of frame generation. Secondly，a pyramid deformation module was utilized to measure the diversity of the generated frames，to calculate the deformation between the generated frames and the original frames to measure the severity of the abnormality. Finally，the anomaly score was obtained by fusing the reconstruction error of the generated frames. The abnormal detection performance of the algorithm was tested on public datasets，and the experimental results showed that the AUC value of the proposed algorithm was higher than that of similar algorithms.

Key words： video anomaly detection； multi-scale；vector quantization； variational autoencoder

0 引言

随着城市化进程的加速和人口密度的增加，公共安全问题日益受到关注。（剩余10816字）