摘要
针对传统视频异常事件检测算法准确率低、鲁棒性差等问题,提出了一种基于双流残差网络的视频异常事件检测算法。该算法综合运用深层残差网络、时序分割网络以及卷积融合策略。在传统双流网络利用单帧图像和多帧光流图像分别提取运动信息和时序行为的基础上,进一步加深网络深度,扩展运动信息建模能力;同时,利用分段构建网络的方式充分提取时序特征,提升对长时间视频处理效果;并且将高维时空特征进行融合,充分挖掘视频中的时空关联关系,得到最终检测结果。在公开的UCF-Crime和XD-Violence数据集上训练和验证的实验结果表明,提出的基于双流残差网络的视频异常事件检测算法相较于仅使用单模态网络(空间流网络)的方法准确率提升约10%,与传统双流网络相比,准确率也分别提升3.2%和6.1%。
To solve the problems of low accuracy and poor robustness of traditional video anomaly event detection algorithms a video anomaly event detection algorithm based on two-stream residual network is proposed.The algorithm uses a combination of deep residual networks temporal segmentation networks and convolutional fusion strategies.Based on the traditional two-stream network the algorithm extracts motion information and temporal behavior from single-frame images and multi-frame optical flow images respectively.The network’s depth is further deepened to extend the motion information modeling capability.The temporal features are fully extracted by using the segmented network construction to enhance the effect of processing long-time videos.High-dimensional spatio-temporal features are fused in the middle layer of the network by convolutional fusion method to fully explore the spatio-temporal correlations in videos and obtain final detection results.Experimental results of training and validation on publicly available UCF-Crime and XD-Violence datasets show that the proposed video anomaly event detection algorithm based on two-stream residual network has approximately 10%improvement in accuracy over methods that only use uni-modal network(spatial stream network).The accuracy is improved by 3.2%and 6.1%respectively in comparison with traditional two-stream networks.
作者
王梓旭
金立左
张珊
苏国伟
陈瑞杰
WANG Zixu;JIN Lizuo;ZHANG Shan;SU Guowei;CHEN Ruijie(Southeast University,Nanjing 210000 China;Xidian University,Xi’an 710000 China)
出处
《电光与控制》
CSCD
北大核心
2022年第8期88-93,共6页
Electronics Optics & Control
关键词
视频异常事件检测
多模态特征融合
残差网络
双流网络
video anomaly event detection
multi-modal feature fusion
residual network
two-stream network
