Research On Pedestrian Tracking And Re-identification Based On Siamese Network

Pedestrian tracking and re-identification technology is an important research direction in the field of computer vision.Pedestrian tracking is to predict the subsequent motion trajectory of a pedestrian based on the state of the pedestrian in the initial frame of the video.Pedestrian reidentification is to measure the similarity of pedestrian images at different points by technical means to determine whether they are the same person.Due to the variable postures of pedestrians and the differences in the environment,pedestrian tracking and re-identification technologies are susceptible to factors such as occlusion,deformation,and out of view,resulting in low accuracy of the algorithm,which makes it difficult to meet the realistic needs in practical applications.Therefore,pedestrian tracking and re-identification methods have certain research value.This paper focuses on the research of pedestrian tracking and re-identification based on the siamese networks.The specific work is as follows:In terms of target tracking,an improved target tracking method based on attention mechanism is proposed for the problem of insufficient feature extraction of SiamRPN method in complex scenes.First,the network -50 is adopted,and a lightweight dual-channel attention mechanism is introduced to optimize feature extraction along the channel and spatial dimensions.Then,a channel attention module is introduced in the bridge connection of ResNet-50 for feature weighting processing.Finally,the generated feature maps are input into the region proposal network for classification and regression to obtain the final tracking results.The experimental results show that the success rate and precision rate on the OTB2015 dataset have increased by 0.5% and 0.7%,respectively.And the accuracy and EAO on the VOT2016 dataset have increased by 2.8% and 0.9%,respectively.In terms of pedestrian tracking,a pedestrian tracking method based on attribute description is proposed to address the problems of low utilization of pedestrian inherent features and serious intra-class interference.First,the attribute appearance information of pedestrians is manually labeled,and the target region mask corresponding to the description information is generated by the attribute description module.Then,SiamRPN is used as the basis of the pedestrian tracking method to generate the target region proposal box of the input image.Finally,the confidence measure and fusion decision are performed on the region mask and region proposal box,and the final pedestrian tracking results are output.The experimental results show that the success rate and precision rate on the OTB2015 dataset have increased by 8.4% and 5.7%,respectively.In terms of pedestrian re-recognition,a pedestrian re-recognition method based on local feature segmentation is proposed to address the problem of low accuracy of pedestrian reidentification.First,the Res Ne Xt-101 network optimized by the IBN-a layer is used for feature extraction.Then,the PCB(Part-Based Convolutional Baseline)network averaging strategy is adopted to slice and segment the generated feature maps to obtain local descriptors.Finally,the joint classification and verification network is used for pedestrian identity classification and similarity determination.The experimental results show that the Rank-1 and m AP on the Market-1501 dataset have increased by 4.8% and 12.1%,respectively.And the Rank-1 and m AP on the Duke MTMC-re ID dataset have increased by 8.1% and 11.9%,respectively.In terms of software implementation,the Pytorch deep learning framework is used to implement the core algorithm in Python language and Matlab R2018 a for interface development to program the pedestrian tracking and re-identification software,which is validated on public datasets and real scenes.The software has the functions of target tracking,pedestrian tracking,and pedestrian re-identification.