改进时空图卷积网络的视频异常检测方法

Improved spatio-temporal graph convolutional networks for video anomaly detection

为了对异常事件中对象的时空相互作用进行精准捕捉,提出一种改进时空图卷积网络的视频异常检测方法。在图卷积网络中引入条件随机场,利用其对帧间特征关联性的影响,对跨帧时空特征之间的相互作用进行建模,以捕捉其上下文关系。在此基础上,以视频段为节点构建空间相似图和时间依赖图,通过二者自适应融合学习视频时空特征,从而提高检测准确性。在UCSD Ped2、ShanghaiTech和IITB-Corridor三个视频异常事件数据集上进行了实验,帧级别AUC值分别达到97.7%、90.4%和86.0%,准确率分别达到96.5%、88.6%和88.0%。

An improved spatio-temporal graph convolutional network for video anomaly detection is proposed to accurately capture the spatio-temporal interactions of objects in anomalous events.The graph convolutional network integrates conditional random fields,effectively modeling the interactions between spatio-temporal features across frames and capturing their contextual relationship by exploiting inter-frame feature correlations.Based on this,a spatial similarity graph and a temporal dependency graph are constructed with video segments as nodes,facilitating the adaptive fusion of the two to learn video spatio-temporal features,thus improving the detection accuracy.Experiments were conducted on three video anomaly event datasets,UCSD Ped2,ShanghaiTech,and IITB-Corridor,yielding frame-level AUC values of 97.7%,90.4%,and 86.0%,respectively,and achieving accuracy rates of 96.5%,88.6%,and 88.0%,respectively.