Research On Person Re-identification Based On Deep Learning

Compared with the target detection technology such as face recognition,the difficulty of pedestrian detection in person re-identification is that pedestrians are all in a free state,which is easy to generate mutual occlusion,ideally,the distance between each sample belonging to a certain category in deep metric learning should be less than the distance between any sample under that category and samples under other categories.However,due to lighting,camera angles,and human movement patterns,The color of clothing,etc.,often cause the distance between the target sample and the positive sample to be greater than the distance between the target sample and the negative sample.The proposal of the traditional triple loss function provides an idea for solving this problem,but The traditional triple loss only considers the distance constraints between the target sample point and the positive sample point and between the target sample and the negative sample sample.This will cause the positive sample sample to be close to the negative sample sample while being close to the target sample.In response to the above problems,the specific work of this article is as follows:(1)This article uses the YOLO(You Only Look Once,YOLO)target detection algorithm for pedestrian detection,and improves the problem that it is easy to miss detection for targets with weak target detection ability and high overlap.In the traditional YOLO algorithm,when a pedestrian candidate box a overlaps with the target candidate box b by more than a given threshold,the algorithm will force the confidence of the candidate box a to zero,making the overlapping target candidate box a impossible It was detected,resulting in missed detection.This paper improved it,and used an improved non-maximum suppression algorithm with linear attenuation to replace the YOLO fixed-threshold non-maximum suppression algorithm.At the time of the box,the method of this paper will attenuate the confidence of the candidate box a,instead of directly zeroing.At the same time,this article draws on the practice of feature engineering in traditional machine learning.A batch normalization layer is added between the convolutional layers of the original YOLO neural network,so that the data has a fixed mean of 0 before convolution.YOLO's network structure speeds up the training of the model.(2)This paper proposes a triple loss function that considers the distance constraint between positive and negative samples to make up for some of the shortcomings of traditional triple loss function in pedestrian distance measurement learning in pedestrian re-identification.The traditional triple loss function does not consider the distance constraint between positive and negative samples,resulting in slow convergence and low accuracy of the neural network.In this paper,the improved triple loss function takes into account the distance between the target sample and the positive sample and the distance between the target sample and the negative sample,and adds the distance constraint between the positive sample and the negative sample.It can make the positive samples close to the target samples and away from the negative samples,so as to better measure the similarity of the samples.