Research And Application Of Person Re-identification Based On Mid-level Features

Intelligent video surveillance has become an inevitable trend in the development of the field,person re-identification is the task of determining whether the pedestrian in a video is the same with one in other videos which are from different camera views in a non-overlapping multi-camera surveillance system.,as the key technology in intelligent video surveillance has become a hot topic in nowadays research.However,due to different camera in front of the body posture and change of light intensity and different camera angles and pedestrian own obstacles,these factors may lead to the same pedestrian different images in the visual apparent difference for pedestrian Recognition brings many difficulties.The current mainstream person re-identification method are often using low-level features,without considering the semantic attributes of pedestrians,and most of these methods using local matching method,without considering the constraints of pedestrians.Aiming at the above problems,this paper makes an in-depth study on the mid-level features and the constraints of pedestrians themselves:(1)A systematic summary of the current person re-identification technology is presented,and the person re-identification is summarized into the following two aspects: "feature design and selection" and "metric learning".And on the basis of the in-depth study of the mid-level features of the person re-identification in the application,including the training of the middle filter process,the mid-level features of the calculation method,and so on.(2)The effect of spatial constraint on person re-identification.We propose a method for person re-identification which fuses the low-level densely patch-matching and the mid-level features with the global vertical appearance constraints,this method considers the two matching layer characteristics of pedestrian images at the same time,the concept of constrained global table is introduced in the vertical direction of the respective mid-level features pedestrian image on the bottom of the fusion of dense block matching method,and validate the algorithm in CUHK01 and VIPeR two public database.(3)Considering the practical problems such as pedestrian detection and the influence of the same pedestrian repeat image in video applications,this paper attempts to design and implement the person re-identification system by using the method of person re-identification which fuses the low-level densely patch-matching and the mid-level features with the global vertical appearance constraints.