Research On Person Re-identification Method Based On Deep Learning

Deep learning is widely used in image recognition,natural language processing,and speech recognition.Pedestrian re-identification is a hot trend in the applications of deep learning technology.Pedestrian re-identification aims to achieve pedestrian matching in a camera network without overlapping surveillance areas.The pictures used by pedestrian recognition are from different cameras,so there maybe great changes in perspective,illumination,size,etc.How to accurately identify target pedestrians under different camera conditions is an important challenge in this thesis.Currently,there are two mainstream research directions in pedestrian recognition: feature representation and metric learning.This thesis combines feature representation and metric learning through deep learning to construct a pedestrian recognition system based on deep learning.The main work and contents are summarized as follows:(1)Based on the deep learning network-RetinaNet,the object detection research is carried out.The structure of the detection network is introduced in detail.The reason for choosing this network structure is that its focal loss function can well alleviate the problem of imbalance between positive and negative samples.In this thesis,the detection network is reproduced through the PASCAL VOC 2012 data set,and the experimental simulation results show that the trained target detection network can detect pedestrian images in complex scenes.(2)Based on the deep learning network-Unet,the image segmentation is studied.The original Unet network will lose some original image information due to the downsampling process,which affects the image segmentation effect.In order to alleviate this effect,this thesis proposes a method of changing the input from a single scale to multi-scale and then sending the output to the corresponding network layer for training.In this thesis,the above method and the original Unet are trained on the same pedestrian data set.The experimental results show that the proposed method has better performance than the original Unet in both quantitative analysis and qualitative analysis.(3)Based on the siamese network,the metric learning algorithm is studied.Because the dataset of pedestrian images is insufficient,this thesis uses pre-training network to alleviate the problem of insufficient data.The mapping function learned by the original siamese network is not enough to make the similar ones closer and the different ones farther.Based on the features extracted from the pre-training network,this thesis proposes a feature combination method to enrich the characteristics of back-end network learning.In this thesis,the test is carried out in the pedestrian recognition dataset Market-1501,and the Rank1 is 85.65% on this dataset.In addition,from the results of the single image,the method proposed in this thesis achieves a higher score for the same categories and a lower score for the different categories.(4)Based on Python's graphical interface design,the interface language used in this thesis is Python.Through its own interface design library Tkinter,the detection,segmentation and recognition functions are combined to form a simple interactive crossrecognition system.