Research And Implementation Of Cross-View Pedestrian Tracking

Cross-field pedestrian tracking refers to the retrieval,matching and tracking of a given target pedestrian in different camera fields of view and has broad application prospects in the field of intelligent security.However,due to the interference of factors such as occlusion,lighting changes and large differences in pedestrian appearance,cross-field pedestrian tracking methods still suffer from poor feature extraction and low tracking accuracy,which make it difficult to meet the requirements of on-the-ground applications.Therefore,the research on cross-view pedestrian tracking is still of great importance.This paper focuses on the cross-view pedestrian tracking method in terms of pedestrian tracking and pedestrian re-identification respectively.The specific work is as follows:To address the problem of insufficient use of image information in pedestrian tracking network models,an improved pedestrian tracking method based on DIMP is proposed.First,the arrangement of residual modules and the short-circuit connection method are changed to optimize the information flow path of the deep residual network.Next,the depth features of the input image are extracted using the improved deep residual network and classified by the convolutional block(Cls Feat).Then,the model predictor processes the features,generates weight information and updates the sample model.Finally,the convolutional layer(Conv)selects the target area based on the weights and the prediction.The target region is selected by the convolution layer(Conv)based on the weights and the confidence of the frame.Experiments on the OTB-2015 dataset showed that the method improved accuracy and success rates by 1.8% and 1.1% respectively compared to the baseline method.To address the problem of small perceptual field of the pedestrian tracking model,an improved pedestrian tracking method based on ATOM is proposed.First,Pyramidal Convolution(Py Conv)is incorporated into the backbone network for capturing information at different levels.Next,Py Conv Res Net is used to extract the image features of the reference and test frames respectively.Then,the target estimation module obtains the final target area by calculating the Io U score and gradient ascent maximization method to obtain the final target region.Finally,the tracking results are obtained by the target classification module.Experiments on the OTB-2015 dataset show that the accuracy and success rate of the method are improved by 1.6% and 1.7% respectively compared to the baseline method.To address the problem of insufficient feature extraction capability of the pedestrian reidentification network model,an improved pedestrian re-identification method based on OSNet is proposed.First,a full-scale residual network incorporating a channel attention mechanism is used to capture features with different sensory field sizes.Then,the feature maps are fed into a unified aggregation gate,which is dynamically fused according to the weight.Finally,the learned features are mapped to a new space using metric learning computation to obtain the rerecognition results.Experiments on the Market1501 and Duke MTMC-Re ID datasets show that the m AP of the method is improved by 1% and 1.1% respectively.In terms of software implementation,the Pytorch deep learning framework is used to implement the core algorithm in Python language and Matlab 2018 b for interface development to program the cross-view pedestrian tracking software.The software has three main functions:pedestrian tracking,pedestrian re-identification and continuous tracking of pedestrians across the field of view,and finally,the video of real scenes is captured for functional verification.