Research On Key Technologies For Pedestrian Tracking In Non-Ideal Environments

Pedestrian target tracking algorithms mainly continuously track target objects based on video or image sequences.Nowadays,pedestrian target tracking algorithms have good tracking performance in ideal environments,but it faces significant challenges in non-ideal environments recognized in the field.Non-ideal environments mainly refer to situations where the target is small,pedestrians block each other,and there is insufficient light,which can greatly affect the accuracy of tracking results.In addition,the performance of detection largely based tracking algorithms currently used is directly affected by the results of target detection.In non-ideal environments,target detection performance is susceptible to interference,which in turn affects tracking results,resulting in issues such as target loss and tracking errors.This article uses YOLOv5＿DeepSort,a deep learning object tracking algorithm,focuses on researching and improving the detection and tracking modules of the target,among them,YOLOv5 is the algorithm detector,and DeepSort is the algorithm tracker,the YOLOv5 algorithm is combined with DeepSort complete tracking tasks,where YOLOv5 is responsible for target detection,while DeepSort tracks the targets detected by YOLOv5,ultimately achieving real-time tracking of multiple pedestrian targets.So as to enhance the algorithm’s performance and increase the dependability of its tracking results in two aspects.This paper mainly focuses on pedestrian tracking in non-ideal environments,and the research content is as follows:(1)A pedestrian detection algorithm based on bidirectional weighted feature pyramid and fused attention is proposed to address detection sissues such as missed detection,false detection,and susceptibility to background element interference in non-ideal environments.A new attention mechanism and feature pyramid fusion method have been introduced into this algorithm,where the attention mechanism can enhance the ability of resource allocation,while the new feature pyramid can improve the ability of feature extraction and fusion.This method achieves better sorting results and improves the accuracy of object detection by replacing the intersection to union ratio function and replacing non maximum suppression methods;The hardwish activation function is used to reduce the "death" rate of neurons and improve the detection accuracy and speed;Combined with the warm-up training mechanism,it reduces learning rate fluctuations,stabilizes learning rate,and improves model training effectiveness.Based on the experimental results of this article,it is shown that the precision rate,speed,and robustness of the model have been improved to some extent in non-ideal environments.(2)Using the above algorithm model as the detector for the target tracking algorithm,a tracking algorithm based on lightweight feature extraction module and multidimensional matching is proposed.This method optimizes the feature extraction model of the tracker for the problems of large model size and low efficiency in tracking multiple targets of pedestrians.It uses separable convolutional control channels to control the number of output channels and lighten the model,thereby optimizing the missed and false detections caused by excessive targets in non-ideal environments,and improving the tracking efficiency and accuracy of the model.In an effort to tackle the challenge which associated with target loss and identity switching that arise from identity information mismatch and target being easily occluded,the corresponding loss function is designed.After the tracking algorithm completes cascade matching,the target position is matched again.By using a complex dataset with a large number of pedestrians to train the model,the effectiveness of this method was verified.According to experimental results,the tracking speed and accuracy of this method have been improved in non ideal environments,and it is more suitable for practical situations.(3)The system uses the aforementioned algorithm to build a deep learning pedestrian multiobjective tracking platform,which tracks a large number of pedestrians in the video and achieves good results.