Research On Pedestrian Re-identification

The purpose of Pedestrian Re-recognition is to match the same person in the video with cross-camera scene. And it has great significance in the realization of intelligent monitoring, cracking down on criminals and other aspects.In this paper, considering the generalization ability and practicability of the existing re-identification system are not strong enough, we research from the aspects of feature extraction, significant learning and metric learning, to propose a pedestrian re-identification algorithm based on convolutional neural network (CNN) convolutional feature weighted block matching and sub-regional metric learning. The main work includes the following three aspects.For the feature extraction, a feature representation method of depth neural network is proposed, which can improve the expression ability of the feature by introducing CNN convolution layer feature in local area.For the weight learning, a significant learning algorithm based on KMEANS clustering is proposed to improve the performance of image block matching. Moreover, it has the advantages of low time complexity and high top k accuracy compared with the existing two methods.For the similarity measure, a localized metric learning algorithm based on dense block is proposed to improve the accuracy of the algorithm. In addition, an improved weighted match similarity formula is proposed to reduce the dependency of the algorithm on the parameters.Experimental results on several standard datasets show that the proposed algorithm can achieve high accuracy. At the same time, the training time is short, the requirement for training set scale is small, and the decline for accuracy rate of cross-dataset training is not obvious,which makes the algorithm has strong practicability.