Research On Pedestrian Detection And Tracking Methods Based On Deep Learning

With the rapid development of science and technology,computer vision technology has also made great progress,especially in robot vision,intelligent driving,security and other aspects,which are inseparable from pedestrian detection and tracking.Pedestrian detection,as the name implies,is a technology to detect pedestrians in images or videos,while pedestrian tracking is a technology to track and predict the trajectory of pedestrians.Pedestrian detection and tracking as the key technology of computer vision,has drawn a lot of constraint conditions,although now have pedestrian detection and tracking applications on the market,but there are a few problems,in particular,environmental factors,occlusion and small-scale pedestrian detection are closely related to the shooting Angle of the camera and environmental factors.In view of these problems,several different detection and tracking methods are proposed in this paper.The specific research contents are as follows:(1)In order to solve the concern of the network on the visible area of pedestrians,a pedestrian detection method combining with the attention mechanism of Center Net is proposed.The use of Anchor in Center Net can effectively improve the detection speed.At the same time,combined with the attention mechanism,some unimportant information can be weakened and important features of pedestrians can be selected,so that it is not easy to lose.This method has a good performance in the experimental results;(2)In order to extract small scale pedestrian target features effectively,this paper proposes a pedestrian detection method based on TRBBC-Net.The TRB basic module is designed to solve the limited ability of single module in traditional Center Net to extract different levels of semantic features and improve the detection effect.In addition,the maximum suppression activation function is introduced to prevent numerical explosion to improve the capability of the detector;(3)For track association in pedestrian process such as the location of the focus on box top information,this paper proposes a pedestrian tracking method based on Center Track,just input a frame before thermal image with pedestrians can predict the location of the next frame pedestrians,testing box is no longer overlap calculation at the same time,but through the pedestrian center ID(Identity document),Improving the tracking effect while reducing the amount of computation suitable for real scene applications.