关键词: 视频异常检测, 深度学习, 人体骨架, 图卷积网络, 注意力

Abstract: Human skeleton has been widely used in the field of behavior recognition, and as a topological structure description method, it has good robustness to light changes and background noise, so it is very suitable for the study of video pedestrian anomaly detection. In recent years, spatiotemporal graph convolutional networks have been used to construct models for detection. However, most of the methods used to describe the strength of human skeleton connection only consider directly connected nodes, focus on small moving areas, and ignore local features. It is still very difficult to accurately detect pedestrian abnormal events. Therefore, a video pedestrian anomaly detection algorithm, PAD-SGMA, based on skeleton graph and mixed attention, is proposed. This method first expands the association between skeleton points, connects the root node with the node that is not directly connected, and divides the human skeleton graph to obtain the local features of the human skeleton. In the graph convolution module, static global skeleton, local region skeleton and attention-based adjacency matrix are used to capture the hierarchical representation. Secondly, a new convolutional network of spatiotemporal channels mixed attention graphs is proposed to increase the attention space and channel relations of the mixed attention module, which helps the model enhance the distinguishing features and give different levels of attention to each joint. In order to verify the proposed model, experiments are conducted on a large-scale open standard dataset (ShanghaiTech Campus dataset), and the results show that PAD-SGMA is more accurate than other methods.

Key words: video anomaly detection, deep learning, human skeleton, graph convolutional network, attention