拓扑信息引导的视频异常行为检测方法

Topology Information Guided Video Abnormal Behavior Detection Method

在视频异常检测任务中,良好的特征提取能力在多帧预测方法中十分重要。然而当面对复杂的环境时,传统的基于空间特征的提取方法往往在多层卷积的过程中忽略了底层特征之间的全局依赖关系。为了更好地进行特征提取,提出一种依托拓扑强相关信息引导的视频异常检测方法。该方法针对底层特征序列进行全局相关性信息的提取,并以此初步增强特征中强关联的信息。将底层特征作为节点,裁剪后的相关性信息作为邻里矩阵,构建关键特征之间的拓扑结构关系图,有效地利用了关键特征的拓扑结构信息。将初步增强的特征与拓扑结构特征进行特征融合,帮助模型更深入更全面地筛选关键特征,提高了特征表达能力。该方法在Ped2、Avenue和ShanghaiTech三个公开数据集上取得了良好的视频帧预测效果,提高了模型的检测精度。

In video anomaly detection tasks,feature extraction capability is very important in multi-frame prediction method.However,in the face of complex environments,traditional spatial feature-based extraction methods often ignore the global dependency relationships between lower-level features in the multi-layer convolution process.Therefore,it is difficult to comprehensively understand the correlation between continuous behaviors in videos.For better extraction results,this paper proposes a video anomaly detection method guided by topological correlation information.Firstly,global correlation information is extracted for the lower-level feature sequence to preliminarily enhance the strongly correlated information in the features.Then,the lower-level features are taken as nodes,and the clipped correlation information is taken as the adjacency matrix to construct a topology structure relationship graph between the key features,which effectively utilizes the topological structure information of the key features.Finally,the preliminarily enhanced features are fused with the topological structure features to help the model more deeply and comprehensively screen the key features and improve the feature expression ability.The method achieves good video frame prediction on three publicly available datasets,Ped2,Avenue and ShanghaiTech,and improves the detection accuracy of the model.