视频动作定位中密集特征金字塔主干网络

Dense Feature Pyramid Backbone Network in Video Action Location

针对视频动作定位算法中金字塔层数增加时间分辨率降低,导致细节特征不完整,进而影响预测结果不准确的问题,提出密集连接型特征金字塔主干网络。视频图像输入特征金字塔主干网络中,密集连接金字塔提取帧级特征和层级特征,实现特征提取阶段参考层、基础层特征与深层特征联系;帧级特征和层级特征通过预测阶段、动作起止时间及标签信息;预测阶段输出融合光流信息输出、动作起止时间及标签预测结果。在THUMOS14数据集的检测结果与AFSD相比,平均精度均值(mAP)提高0.4%,准确定位动作在视频中的起止时间和类别,可应用于智能监控等场景。

The number of pyramid layers increases and the time resolution decreases in the video action location algorithm,resulting in incomplete detail features,which affects the accuracy of prediction results.Aiming at the problem,this paper proposes a densely connected feature pyramid backbone network.The video image is input into the backbone network of the feature pyramid,and the pyramid is densely connected to extract the frame level features and hierarchical features,so as to realize the connection between the reference layer,foundation layer features and deep features in the feature extraction stage.Frame level and hierarchical features output start and end time of actions and label information in the prediction stage,and the fused optical flow information and label prediction results are also output.On the THUMOS14 dataset,compared with AFSD,the mean average precision(mAP)is improved by 0.4%.It can accurately locate the start and end time and category of actions in the video,and can be applied to intelligent monitoring and other scenes.