Research On Person Re-identification Based On Feature Learning

Person re-identifction refers to the process of matching the same pedestrian image that appears in cameraswithout overlapping horizons.In recent years,the paper related to person Re-identifction has been growing rapidly.This technology has become an important branch of intelligent video surveillance system.Person re-identifction is currently facing challenges such as light,gesture,perspective and other changes.At present,the related research of person re-identification has the tendency of extracting features from manual to automatic learning,from image-based re-identifction to video-based re-identifction,from closed re-identifction to open re-identification.The basic steps of person Re-identifction are divided into two basic steps: feature extraction and metric learning.This paper designs a deep network FT-FFN(Fine-tunning Feature Fusion Network)for extracting features.On one hand,the network fuses CNN features with low-level visual features by means of a fully connected layer.In the process of identifying the pedestrian attribute in the process of network training,the low-level visual feature in the process of back-propagation and updating the weight will affect the adjustment of the CNN network part,so that the CNN feature changes in the direction complementary to the low-level visual feature.On the other hand,in the process of fine-tuning with the identification of pedestrian attributes,the distinguishing ability of network fusion layer output feature is greatly enhanced,and finally the fusion layer feature of network output has excellent distinguishing ability.In addition,the output layer of FT-FFN generates the semantic attributes of pedestrians.In this paper,we discuss the types,frequencies and distinguishing abilities of semantic attributes,improve the optimization method of middle attribute weight,and combine attribute featurewith fusion layer in the person re-identifction.In this paper,the experiment and test of the semantic attribute of the middle layer are carried out on the PETA dataset.The experiment design and test of the VIPeR,PRID 450 s and CUHK01 datasets are carried out.The results show that the FT-FFN is more accurate in the semantic attribute,FT-FFN fusion feature is discriminative.