Deep Learning Approach For Pedestrian Re-identification

Pedestrian re-identification is an important technology in computer vision.It has broad application prospects and research value in the fields of video surveillance and intelligent video analysis.Among them,the pedestrian image comes from different monitoring.Under the different monitoring of the same pedestrian,due to the change of illumination,the change of the angle of view,the difference of the posture,etc.,it will cause a big difference,even under different cameras,different pedestrians.It may be more like the appearance of the same pedestrian,which makes pedestrian recognition a very challenging topic.Therefore,the research focus of pedestrian recognition is to find discriminative features and metrics that can be stable under cross-view.In this paper,from the image itself,the dominant image background is redundantly eliminated by the saliency detection algorithm.The foreground of the characters is extracted and the features are extracted.On the other hand,with the hot learning of deep learning,the convolutional neural network is widely used for its powerful feature extraction ability.we constructs a neural network to extract robust discriminative features.And then identify.The specific work of this paper is as follows:(1)Aiming at the background redundancy problem in different perspectives of pedestrian recognition,a saliency detection preprocessing algorithm based on simple linear iterative segmentation is designed.It is mainly used to eliminate the redundant background of the image background before feature extraction,so that the feature extraction can obtain more robust features and improve the recognition rate.(2)A pedestrian re-identification algorithm based on the block-based verification network model is designed.Each pair of input pedestrian sample pairs is divided into three pairs of head,body and foot,which are input into three pairs of siamese networks for training.This network of block structure facilitates the network to perform better feature extraction on the image part,and then comprehensively judges the results of the three parts of the recognition,so that the result is more accurate.(3)Based on the verification network model proposed in the previous chapter,the verification and verification network model and the identification of the network model use the tags,and each sub-network part is changed into a separate authentication network,so as to fully utilize the similarity tags and ID tags.At the same time,for the problem of poor recognition on small data sets,this chapter uses the feature of generating a confrontation network to convert the data set styles to each other,and increases the number of samples in the original data set,thereby improving the recognition rate.