Research Of Video Person Re-Identification Based On Graph Convolutional Network

Person re-recognition technology is an important task in video surveillance.When the camera cannot obtain high-quality face images,it uses more comprehensive information such as body characteristics and gait movements to continue to identify the target,whether it is used alone or with a person.The combination of face recognition technology can play an important role and has important academic research significance in the field of computer vision.At present,person re-recognition based on single-frame images has made great progress,but the image information is limited and there are greater requirements for image quality.Considering that the video-based person re-recognition method not only focuses on the information of single frame image,but also can utilize the time sequence information between frames.In response to the challenging problems of the occlusion and the visual ambiguity in re-identification task,the main work and research results of this paper include:1)Propose a video person re-identification method based on person attribute graph convolutional network.Inspired by the multi-task network,considering that person attributes can provide more fine-grained features to capture the discriminative information between different person,so the attribute features are integrated into the re-identification task.The algorithm is divided into person attribute prediction branch and person re-identification branch.The person re-recognition branch combines the global features with the attribute features obtained by the attribute prediction branch to obtain a more comprehensive person feature,and then uses the graph convolutional network to learn the potential connections between different video frames,and extracts the robust discriminative features to solve challenging problems such as occlusion and visual similarity in re-identification tasks.The experimental analysis on the dataset MARS shows that the proposed algorithm has achieved good recognition results and can effectively improve the recognition performance.2)Propose an improved video person re-identification method based on person attribute graph convolutional network.Aiming at the lack of certain attribute information in partial video frames,the spatial attention mechanism and channel attention mechanism are introduced to mining notable local attribute features,and capturing the subtle distinguishing features between similar person to improve the re-recognition algorithm.Experiments on the dataset MARS verify that the recognition accuracy and effectiveness of the improved algorithm have been improved.