Research On Improved Algorithm Of Person Re-identification Based On Deep Learning

The research of pedestrian re-identification technology can greatly improve the low efficiency of the current monitoring system and make the implementation of intelligent security systems possible.Although pedestrian re-recognition technology has become a hotspot in the field of CV in recent years and many research results have been achieved,most of the existing algorithms use multi-branch structure to solve the problem of misalignment between images,which brings a lot of calculations.This topic uses a single branch structure in the network structure to take advantage of deep learning to extract features automatically,while optimizing the metric loss function to achieve the purpose of reducing the amount of network calculations and making the model more stable.The main work done in this paper is as follows:Firstly,this paper fully studies the effective tricks for pedestrian re-identification.Good training skills play a vital role in improving model performance,this topic designs a basic network structure based on the SE?Res Net50,which is called Backbone.Based on the Backbone,this article does ablation experiments for the tricks algorithm commonly used,and verifies the influence of each algorithm,and then obtains a New Baseline.Secondly,a Multi-stage Attention structure based on the attention mechanism is designed.For feature comparison,the existing algorithms mostly divide the picture in a certain way compulsorily(with the help of segmentation network,pose estimation network,etc.),and use a multi-branch structure to match pedestrian features,while the multi-branch structure increase the amount of calculation greatly.This paper studies the attention mechanism in depth and applies it to the pedestrian re-recognition technology,and designs the Multi-stage Attention structure so that the network can automatically focus on different areas according to the importance and extract more distinguishing features.The Multi-stage Attention structure solves the problem of feature matching without additional calculation.Then,the metric loss function is optimized to make that the network can extract more distinguishing features and make the training process more stable.This topic studies the loss functions commonly used in detail.Among them,the triplet loss suffers from difficulties of selecting sample pairs,and time-consuming during training.Center loss can optimize the distance within a class,but there is no constraint on the distance between classes.This paper improves the measurement loss function based on triple loss and center loss,and mines difficult samples around the center of the class.It avoids the phenomenon that the model is not easy to converge stably due to the randomness of the sample pair distribution when searching for difficult samples by exhaustive sample pairs.Finally,combining with effective training techniques for Re-ID and the Multi-stage Attention module and the improved metric loss function,and improving the accuracy of the pedestrian re-identification algorithm on the public datasets,while increasing the stability of the model.