【摘要】 视频异常行为检测是目前计算机视觉领域的热点问题之一。然而,由于异常行为难以具体定义,使得基于监督学习的二类分类方法难以应用在该领域。本文提出了一种无监督的视频异常检测模型称之为基于时空特征融合的3D自编码器模型(ST-3DCAE)。模型采用PWCNet提取场景光流特征图,并与原视频帧融合作为基本单元,由多个基本单元组成连续基本单元作为模型的输入;利用3DConv和ConvLSTM模块进行时空特征的自主提取,3DSEblock模块进行重要特征的筛选;最终通过输入数据和自编码器重建视频块之间的重建误差,来判断视频是否出现异常行为。通过在UCSD、Avenue等公开数据集上进行验证实验结果的定性和定量分析证明了本方法具有较好的性能。更多还原

【Abstract】 Video anomalous behavior detection is one of the current hot problems in the field of computer vision. However,the difficulty of specific definition of anomalous behavior makes it difficult to apply supervised learning-based two-class classification methods in this field. In this paper,we propose an unsupervised video anomaly detection model,called 3 D self-encoder model based on spatio-temporal feature fusion( ST-3 DCAE). First,the optical flow feature map of the scene is extracted with PWCNet and fused with the original video frame as the basic unit,and multiple basic units form a continuous basic unit as the model input;then,the 3 DConv module and ConvLSTM module are used for autonomous extraction of spatio-temporal features,and the3 DSEblock module is used for screening of important features; finally,the reconstruction of the self-encoder through the input data and reconstruction error between video blocks to determine whether the video shows abnormal behavior. The model proposed in this paper was validated on publicly available datasets such as UCSD and Avenue,and the qualitative and quantitative analysis of the experimental results proved the better performance of this method.更多还原