Aiming at the problems that the complex geometry of moving targets makes edge fea-ture extraction insufficient and leads to inaccurate edge detail positioning,a position attention network(PA-Net)is proposed.By introducing a position attention network on a feature map of moderate size,according to the significant difference between the moving target edge and the background pixels,horizontal and vertical attention weights are dynamically assigned to different positions of the target,and the degree of attention to the target edge is enhanced.It not only captures the dependency between the foreground and the background,but also retains the exact position information of the target,thereby enhancing the ability to extract the details of the target edge and improving the accuracy of target framing.The experimental results show that compared with the benchmark video object detection network StreamYOLO,the detection results of the PA-Net algorithm on the Argoverse-HD dataset show an average improvement of 0.3%,1.6%,and 0.6%in detection accuracy under different intersection over union,respectively,which has certain application prospects in the fields of robotics and autonomous driving.
关键词
视频目标检测/目标定位/位置注意力/边缘特征
Key words
video object detection/object location/position attention/edge features
维普
万方数据