Video Anomaly Detection Based on Space-Time Fusion Graph Network Learning

There are strong correlations among spatial-temporal features of abnormal events in videos. Aiming at the problem of performance for abnormal event detection caused by these correlations, a video anomaly detection method based on space-time fusion graph network learning is proposed. In this method, spatial similarity graph and temporal trend graph for the segments are constructed in terms of the features of the segments. The spatial similarity graph is built dynamically by treating the features of the video segments as the vertexes in graph. In this graph, the weights of edges are dynamically formed by taking the relationship between vertex and its Top-k similarity vertexes into account. The temporal trend graph is built by taking the time distance for m sequential segments into account. The space-time fusion graph convolutional network is constructed by adaptively weighting the spatial similarity graph and temporal trend graph. The video embedding features are learnt and generated by using this graph convolutional network. A graph sparse regularization is added to the ranking loss, in order to reduce the over-smoothing effect of graph model and improve detection performance. The experiments are conducted on two challenging video datasets: UCF-Crime(University of Central Florida crime dataset) and ShanghaiTech. ROC(receiver operating characteristic curve) and AUC (area under curve) are taken as performance metrics. Our method obtains the AUC score of 80.76% rising by 5.35% compared with the baseline on UCF-Crime dataset, and also gets the score of 89.88% rising by 5.44% compared with SOTA(state of the art) weakly supervised algorithm on ShanghaiTech. The experimental results show that our proposed method can improve the performance of video abnormal event detection effectively.