Pedestrain Re-Identification Based On Scenes Transition And Region Alignment

Pedestrian re-identification technology is a key technology in intelligent video surveillance.It studies pedestrian recognition and tracking across cameras and retrieves whether it appears in other scenes based on specific pedestrian images.Due to the different time and location of the images captured by the camera,the acquired pedestrian images inevitably have changes in background,perspective,illumination and the view of pedestrians,which leads to great differences among different pedestrians in different monitoring scenarios.It makes the pedestrian re-identification problem facing great challenges.This article mainly studies the pedestrian re-identification based on scene transition and region alignment.The main research contents are:Firstly,due to the fact that pedestrians detection bounding boxes do not fit the pedestrian in the dataset,a sliding window alignment method based on semantic segmentation(SWA)is used to solve this problem.This kind of problem is mainly caused by the excessive background area of the dataset sample and the interference of background objects,which will affect the extraction and recognition of pedestrian.The method uses the foreground area extracted by Refine Net and Mask RCNN respectively for misalignment detection,and partitions the image in the vertical direction to obtain the position of the pedestrian in the image.Using the sliding window mechanism,it can make the pedestrian closely fit the pedestrian detection bounding boxes.Secondly,in view of the current shortage of existing pedestrian re-identification training samples and the low recognition rate,a scene transition method based on pedestrian feature recovery(PFRGAN)is proposed.The foreground pedestrian area is identified by the semantic segmentation method(Mask and RCNN Refine Net),and the image style transition between the scenes is performed by using Generative Adversarial Networks for the background area part,thereby generating auxiliary data with more background changes while retaining the pedestrian features.Thirdly,based on the optimization of the dataset,the residual convolutional neural network structure(Res Net-50)is improved and named double-path augmentation net(DFANet).The global feature branch is added to the structure,so that the deep learning model is more suitable for the aligned pedestrian samples.DFANet also use the Drop Block module.This module discard area with semantic information in the feature map,and increase the weight of other areas in the recognition,and reduce the over-fitting of DFANet.The use of the label smoothing mechanism(LSR)for the generated supplemental data suppresses the difference between the positive and negative samples in the output,thereby improving the adaptability of the deep learning models.Finally,the method proposed in this paper was used in two datasets,named Market-1501 and Duke-MTMC-re ID.The innovations proposed in this paper are compared with the state-of-art re-Id methods.And the Rank-1 indicators achieve 92.2% and 83.4% accuracy on the two datasets respectively.