A Methodology For Person Re-identification In Surveillance System

Non-overlapping surveillance system is widely used in various public places, for public areas with large daily pedestrian volume, such as: shopping malls, parks, train stations and airports, safety monitoring is very important. In non-overlapping surveillance system, there are several complex and uncontrollable situations that increased the difficulty of pedestrian re-identification, including perspective variation, illumination variation, complex background and occlusion. This paper study and research pedestrian re-identification problem based on existing pedestrian re-identification research background, and further explore pedestrian re-identification problem under different scenarios, which conducted the following research:First, review the summary of predecessors' research status briefly. These approaches can be roughly divided into two types, one is based on representative pedestrian feature selection, and the other one is based on metric learning. A variety of methods' mathematical definition and algorithm procedure are described for each type separately, and the more effective methods are explained in detail.Then, according to previous analysis, the existing algorithms that applied better can combine feature selection and metric learning this two types of approaches. On this basis, set-to-set distance computing methods are researched. By comparing several different methods, the most representative method is chosen. Further, some feature descriptors and the learnt metric are combined to form pedestrian re-identification the algorithm based on set-to-set metric learning.Finally, the proposed method is a new pedestrian re-identification method in surveillance system with pedestrian sequence as processing element, and pedestrian sequence is also a set. In this method, RDC is chosen as metric learning algorithm, and set-to-set distance comparison is also used. This new algorithm transform pedestrian re-identification problem into set-to-set metric learning problem with feature selection. The approach proposed in this paper is compared with several state-of-the-art algorithms on NLPR_MCT dataset and i-LIDS MCTS dataset. Experimental results show that this algorithm has a more superior performance than the state-of-the-art algorithms, which improved matching rate and removed mismatch. This algorithm also can be generalized.