Research On Pedestrian Multi-target Tracking Algorithm Based On Deep Learning

With the continuous deepening of deep learning technology and the continuous research and development of all walks of life,pedestrian detection and tracking technology based on deep learning has attracted more and more attention from domestic and foreign university laboratories.Universities and research institutes in various countries have launched research on this Research vigorously.As one of the popular research directions in the field of computer vision technology,object tracking is widely used in the fields of intelligent monitoring,intelligent transportation,human-computer interaction and automatic driving.However,in the actual tracking application,the large amount of parameters based on the deep learning algorithm affects the tracking efficiency,the constantly changing scene environment and the irregular movement of pedestrians in reality make the process of pedestrian tracking vulnerable to illumination and mutual occlusion between targets.And the interference of the change of pedestrian motion scale seriously affects the tracking performance of pedestrians,which is easy to cause problems such as target ID switching and accuracy reduction.Therefore,this paper conducts in-depth research on pedestrian multi-target detection and tracking in complex scenes.The specific research contents are as follows:(1)In view of the large amount of parameters of the deep learning model and the slow tracking speed,this paper performs sparse training on the YOLOv4 detection network and performs model pruning to obtain a lightweight detection model.The detection rate of the improved model has been significantly improved.(2)Aiming at the impact of the detection performance after the model is lightweight and the redundant information will be generated after the up-down sampling and fusion of the YOLOv4 model,the attention mechanism CBAM module is added after the Concat processing to enhance the useful information in the feature information of different granularities after fusion and suppress the redundant information.Other non-important information to improve model accuracy.At the same time,the 1×1 convolution on the PANet layer is replaced by a resnet residual network,which effectively solves the problem of model training gradient explosion.(3)Considering that the Kalman filter in the DeepSORT tracking model is only applicable to the linear model,and pedestrians will have nonlinear motion during the motion process,this paper introduces a nonlinear algorithm to extend the Kalman filter to perform linear fitting on the nonlinear motion;Because pedestrians are easily affected by lighting and occlusion between targets will affect the tracking performance of the model,adding HSV color features can reduce the target ID-switch problem caused by lighting and occlusion to a certain extent.