Research On Person Re-identification Algorithm Based On Deep Learning Of Multi-granularity Features

Person re-identification is a popular computer vision algorithm,which uses images captured by surveillance cameras in life to realize person cross-region retrieval again,and thus person cross-border tracking.The study of person re-identification algorithms is of great significance to the development of public security.Due to environmental constraints,person re-identification is faced with some difficulties in practical applications,such as: pedestrian misalignment,background interference,etc.In order to improve the accuracy of person re-identification and solve the problems of misalignment and background interference in person re-identification,this research improves on the basis of commonly used horizontal segmentation methods,focusing on person re-identification based on multi-grain feature deep learning Identification algorithm.The main research contents of this dissertation are as follows:Research the local area localization algorithm of human body to solve the detection of upper and lower body of human body.In order to improve the success rate of the algorithm for the detection of the upper and lower body of pedestrian images,eliminate the impact of human body structured deformation on the detection effect,and at the same time improve the detection effect of the network on blurred pictures,this algorithm extracts features based on the Yolov1 target detection algorithm The network is replaced,the grid segmentation strategy is modified,the network input image size is modified,and a data preprocessing mechanism is added.Through the experimental test to improve the detection effect of the front and back Yolov1 algorithm on the upper and lower body of the human body,it is proved that the local region positioning algorithm studied has a better detection effect on the upper and lower body of the human body than Yolov1.In order to alleviate the problem of unevenness and background interference of pedestrians,the traditional horizontal segmentation is improved,and a precise segmentation method of pedestrian feature maps is proposed.This method uses the coordinates of the upper and lower body of the pedestrian extracted by the local area localization algorithm to segment the feature map extracted from the backbone network,and re-integrates the corresponding area of the pedestrian to reduce the impact of pedestrian misalignment and background interference on the feature.In order to further improve the accuracy of person re-identification,a multi-granularity feature fusion algorithm is designed,which uses human body features as coarse-grained features,adopts precise segmentation of pedestrian local area features as fine-grained features,and uses mid-level network features as mid-level features.By integrating the multi-granularity information of pedestrian space and the multi-granularity information of network space features,the pedestrian characteristics are enriched and the recognition accuracy is improved.At the same time,when calculating the loss function,different weight values are given for different granularity features,changing the network's attention to different features.The experimental environment was built under the Pytorch deep learning framework,and the effectiveness of the multi-granularity feature fusion algorithm was verified through comparison experiments,the optimal distance calculation function was selected,and the effectiveness of the segmentation network was verified.The results show that on the Market1501 dataset,the proposed multi-granular feature fusion algorithm has initially solved the effect of pedestrian misalignment and background interference on recognition accuracy.