Person Re-identification Based On Deep Learning

In recent years,with the rapid development of artificial intelligence technolo'gy and the continuous advancement of the construction of smart cities,a huge monitoring system composed of surveillance cameras has become an important part of smart cities.Video surveillance system is an important auxiliary measure for urban security prevention and control.How to realize pedestrian tracking and pedestrian retrieval between non-overlapping camera perspectives is particularly important Person Re-identification(ReID)technology is exactly to achieve pedestrian retrieval across cameras,but the task of pedestrian re-identification is still a challenging subject due to lighting,occlusion,resolution and other issues.Based on deep learning algorithms,this article focuses on the following three aspects:1.Aiming at the misalignment of pedestrian images,this paper proposes a pose-guided multi-granularity feature fusion ReID algorithm.Use the PAF network to capture the key points of the human body,the key points of the human body are used to align the corresponding parts between the images.Extract the multi-granularity features after the aligned images,and perform similarity matching to realize the ReID task.This algorithm has achieved good performance only under the ID loss.2.Aiming at the problem that the pose network consumes a lot of resources,this paper proposes a multi-granularity feature fusion ReID algorithm.Different pooling strategies are used to obtain global features and local features.The cosine distance between local features is used to represent human structural features.A multi-level loss supervision mechanism is used to supervise the network and improve network performance.3.Aiming at the problems of insufficient pedestrian datasets and large intra-class differences,this paper proposes a ReID algorithm based on GAN network hybrid coding.The network can achieve controllable image generation by cross-mixing pose features and appearance features in two images.The appearance characteristics of the generated image are fedback to the generation network to realize the online interactive loop and joint optimization.The RelD algorithm proposed in this paper solves the problems of misalignment and data sparseness in pedestrian images.The experimental results in the public dataset show that the proposed algorithm has certain advantages.