A Research On Pedestrian Re-identification Based On Deep Learning

Pedestrian re-identification technology is simply defined as a computer vision technology which can be used to determine whether the pedestrian under the absence of crossing monitoring camera system can be correctly matched.Moreover,this technology combine with pedestrian detection and pedestrian tracking technology,and they are widely used in intelligent video surveillance,intelligent security,intelligent transportation etc.However,due to the shooting angle and the resolution are different,and pedestrians dressing change or movement posture change.And it also met the influence of natural conditions such as rain and snow,flog and illumination.All of those factors make the recognition rate of pedestrian re-identification are low.At present,the mainstream of pedestrian re-identification research focuses on feature representation and distance metric learning.The former methods often need to use effective pedestrian appearance characteristics to extract the distinguishing features and a robustness characteristic on the angle different.And the latter need large training images to learn suitable distance function to measure the similarity between the training image pairs,and this method is not robust to the images with complex heterogeneous distribution.For the above analysis,this essay proposes three optimization schemes based on how to extract effective pedestrian features.Our main contributions to this essay are as following: 1.An improved Siamese network model with feature enhancement capability and a positive sample sampling method are proposed.And this scheme aims to solve the problem that the features extracted by CNN are difficult to distinguish between the positive pairs and negative pairs.Firstly,do pixel-level superposition operations on the extracted features.If there are regions with the same characteristics between the image pairs,the pixels in the region are superimposed and enhanced,otherwise,they are weakened.Secondly,the positive sampling method is used to search and match the data in the dataset dynamically and to find the closest image to the target image and then form them to the training image pairs.This method reduces the difference between image's feature maps by reducing the difference between the image pairs.And experiments show that the improved method can obtain better feature enhancement effect and recognition rate of pedestrian re-identification.2.A pedestrian re-identification method that combines classification and verification models is proposed.This method aims to improve the problem that the previous verification model for pedestrian re-identification has a relatively simple function.Based on the above mentioned Siamese network,the improved Siamese network is further expanded by combining the classification and verification,and uses different loss functions to classify and verify the input person images according to different models.And experiments show that the improved scheme can not only perform classify and verify tasks,but also improve the recognition rate of pedestrian re-identification.3.A pedestrian recognition method based on multi-scale feature fusion is proposed.This method improves the problem that the features extracted by common CNN networks is utilizated insufficient of Single-Scale images.The Multi-scale feature fusion network extracts features from different scales images,and then mixes these features.And then we use L2 to calculate the distance between the mixed features,and the similarity between image pairs can be measured according to the feature's distance.In addition,the mixed features are used to perform Multi-query tasks.Experiments show that the improved method can improve the recognition rate of pedestrian re-identification.