Research On Pedestrian Tracking Based On Deep Learning

Pedestrian tracking research has a wide range of application values in the fields of epidemic prevention and control,autonomous driving,public security and security,and intelligent robots.However,the actual application environment is often complex.Whether it is single-target or multi-target tracking,problems such as target occlusion,small target tracking,and illumination changes will be involved,resulting in a decrease in the accuracy of target tracking.Therefore,improving the accuracy of pedestrian tracking algorithms has research value and challenges.Usually,the main process of pedestrian tracking can be divided into the following steps: first,the deep learning algorithm is applied to detect the pedestrians appearing in the video stream,and the pedestrians in the video are marked;then the pedestrian features in each target are extracted,Secondly,the correlation algorithm is used for data association;finally,the pedestrian targets with the same characteristics calculated by the correlation algorithm are marked with the same ID.According to the complete pedestrian tracking process mentioned above,the main research work in this paper includes the following aspects:(1)The improved YOLOv4 algorithm is selected in the target detection stage.First,reduce the convolution level of the backbone network to improve the speed of detection.Secondly,the DenseNet network is added to the original backbone feature extraction network.By introducing the DenseNet network,the network level is deepened to improve the detection accuracy.Finally,the original activation function Mish of YOLOv4 is changed to Hard-Swish,which reduces the amount of calculation and improves the detection speed of the model.(2)The target tracking algorithm adopts the improved Deep Sort algorithm.First,data enhancement is introduced,by blurring the image extracted by YOLOv4,and adding the ellipse detection frame algorithm.Secondly,the appearance feature extraction network is replaced by Darknet53,and a weakening processing module is added.Through the improvement of the algorithm,the accuracy of data association is improved,thereby improving the accuracy of target occlusion and small target tracking.(3)Design the algorithm verification experiment.Based on the MOT16 dataset with more than 7000 pedestrian images,the experimental results show that the improved algorithm proposed in this paper has improved the accuracy of pedestrian tracking.