基于全局-局部自注意力网络的视频异常检测方法

Novel video anomaly detection method based on global-local self-attention network

为提升视频异常检测精度,提出一种基于全局-局部自注意力网络的视频异常检测方法。首先,融合视频序列与其对应的RGB序列凸显物体的运动变化;其次,通过膨胀卷积层捕获视频序列在局部区域的时序相关性,并利用自注意力网络计算视频全局时序的依赖性,同时,依靠增加基础网络U-Net的深度并结合相关运动和表征约束对网络模型进行端到端的训练学习,从而提升模型的检测精度和鲁棒性;最后,对公开数据集UCSD Ped2、CUHK Avenue和ShanghaiTech进行测试并对所得结果进行可视化分析。实验结果表明,所提方法的检测精度AUC值分别达到了97.4%、86.8%和73.2%,其性能明显优于对比方法。

In order to improve the accuracy of video anomaly detection,a novel video anomaly detection method based on global-local self-attention network was proposed.Firstly,the video sequence and the corresponding RGB sequence were fused to highlight the motion change of the object.Secondly,the temporal correlation of the video sequence in the local area was captured by the expansion convolution layer,along with the self-attention network was utilized to compute the global temporal dependencies of the video sequence.Meanwhile,by deepening the basic network U-Net and combining the relevant motion and representation constraints,the network model was trained end-to-end to improve the detection accuracy and robustness of the model.Finally,experiments were carried out on the public data sets UCSD Ped2,CUHK Avenue and ShanghaiTech,as well as the test results were visually analyzed.The experimental results show that the detection accuracy AUC of the proposed method reaches 97.4%,86.8%and 73.2%respectively,which is obviously better than that of the compared methods.