Research On Multi-camera Pedestrian Detection And Person Re-identification Method

With the improvement of the security monitoring system and the increasing number of video surveillance devices,people are increasingly demanding the intelligentization of video surveillance technology.As the core link in video surveillance technology,person re-identification task uses image processing,computer vision,pattern recognition and machine learning to solve the person search task across cameras.Most existing person re-identification methods are based on the cropped pedestrian image.However,in the actual monitoring scene,pedestrian labeling does not exist.It is necessary to detect all pedestrians in the original video before re-identification.An algorithm framework is proposed,and consists of two independent training algorithms,which are pedestrian detection algorithm and person re-recognition algorithm respectively.Firstly,this paper summarizes the research background,research significance and research status of pedestrian detection and person re-identification technology.Then introduce the theoretical knowledge of the related convolutional neural network.Secondly,in order to improve the accuracy of the pedestrian detection model,this paper proposes a Binary-SSD pedestrian detection algorithm based on the SSD detection algorithm in the multi-target detection field.First,according to the pedestrian aspect ratio,a priori box with a fixed aspect ratio is designed to replace the prior frame with multiple aspect ratios in the original algorithm.Secondly,using multiple data enhancement methods to increase the number of training set samples increases network learning ability.Experiments show that the improved B-SSD pedestrian detection algorithm has improved in accuracy and speed.Then,in order to improve the accuracy of person re-identification model,this paper proposes a person re-identification method based on joint learning and re-ranking.Firstly,in terms of feature description,this paper analyzes and contrasts two models commonly used in person re-identification: verification model and recognition model.In order to make full use of supervision information,this paper combines two models to constrain the network together,so that the network can learn more discriminative pedestrian characteristics.In terms of distance metric,this paper introduces the k-reciprocal nearest neighbor method to reorder the results of the initial distance calculation,and removes the mismatch,so that the more similar correct matching ranks higher,and the re-identification accuracy is improved.Experiments show that the improved person re-identification algorithm has good performance.Finally,the main research contents of this paper are summarized,the shortcomings of the algorithm are pointed out,and the future research direction is prospected.