MPAT: multi-path attention temporal method for video anomaly detection

Hongjun Li, Xiaohu Sun, Chaobo Li, Xulin Shen, Jinyi Chen, Junjie Chen, Zhengguang Xie

Abstract: Video anomaly detection is a recent focus of computer vision research thanks to the rarity and uncertainty of anomalous events. However, most existing research works are limited to learning the apparent and motion information of specific objects, ignoring the effect of temporal information. In this paper, multi-path attentional temporal method is proposed to detect whether videos contain anomalies. Specifically, the activity of adjacent units is regulated by a novel intra-layer Recurrent Residual Convolution Unit (RRCU) with temporal function, and different time steps are set to enhance the model’s ability to integrate contextual information. Furthermore, considering the information loss caused by image compression in the encoding stage, Skip Attention Gates (SAG) are used to focus on specific objects of different shapes and sizes and aggregate information from multiple feature scales. As an end-to-end learning framework, the proposed model can extract more discriminative spatio-temporal features, and the experimental results on three datasets demonstrate the effectiveness and generalization of the proposed approach.