Research On Key Technologies Of RGB-IR Cross-Modality Person Re-Identification

Person re-identification has become an important task in the computer vision.In recent years,with the rapid expansion of the demand for all-day intelligent monitoring and the background of the COVID-19 raging around the world,there is an urgent need for a method that can intelligently process the infrared image data and link it with the RGB visible data to realize cross-mode dynamic pedestrian re-identification.In this paper,we had deeply studied the problem of RGB-IR cross-modality pedestrian reidentification from three perspectives:metric learning,representation learning,and cross-domain adaptation.The main research contents are as follows:1.In the task of cross-modality person re-identification based on metric learning,a new channel joint learning strategy is proposed,and the weighted triplet loss function is improved,using single-channel data as an auxiliary modality improves the effect of RGB-IR cross-modality metric learning.2.In the task of cross-modality person re-identification based on representation learning,a neural architecture search strategy based on BN layer separation selection is proposed,and automatically search for the optimal backbone network architecture suitable for solving the problem,which greatly saves the cost of manually designing the network architecture.3.In the task of domain-adaptive cross-modality person reidentification,a progressive domain-adaptive learning framework based on teacher-student model is proposed,the classification loss is improved,and the impact of clustering noise on model training is further weakened by the dynamic and symmetric cross-entropy loss function,and the cross-domain adaptive performance of cross-modality person re-identification is improved.The algorithms were compared with the current state-of-the-art algorithms on multiple public datasets,the experimental results verify the effectiveness of the proposed algorithms.