Abnormal event detection in surveillance videos based on multi-scale feature and channel-wise attention mechanism

Abnormal event detection is a challenging task, due to object scale variation, impact of background and anomaly defined differently in different context. In this paper, we propose a new multi-scale feature prediction framework for abnormal event detection. Firstly, we construct a multi-scale alignment feature generator to fuse the characteristic of different receptive fields so that address the objects of different scales in video frame. Secondly, in order to weak the influence of background, a novel channel-wise attention mechanism is introduced to highlight those informative channels while suppressing the confusing ones. Finally, an autoencoder-based deep feature prediction module is applied to capture temporal information and contextual information to generate predicted features. Instead of giving a definition of anomaly, we treat predicted features that differ from the actual features as abnormal features. Experimental results on four benchmark datasets demonstrate the superiority of the proposed framework over the state-of-the-art approaches.