Person re-identification algorithm based on multi-module convolutional neural network

Abstract

For the cross-border tracking needs of target persons in real complex scenes, a person re-identification algorithm based on a multi-module convolution neural network is proposed to solve the problem of person search and matching caused by person scale change, light change, posture change and other factors in the real environment. The algorithm takes ResNet50 as the backbone network of feature extraction. The STN network module is embedded into the backbone network to overcome the impact of person scale change. The IBN network module is integrated for person image color correction to compensate for the influence of illumination change in the real scene. And A person multi-branch feature extraction module is designed to effectively reduce the impact of person posture changes. Through person image feature expression and measurement learning calculation, person similarity matching across cameras is realized. Experiments show that this method has good performance in real complex scene test data, and its Rank-1 and mAP are 98.30% and 95.78% respectively. It can be used for person matching and search in a real complex environment, and has certain practical value.