Research On Key Issues Of Person Re-identification In Non-overlapping Multi-camera System

With the extensive application of video surveillance systems,the requirement of intelligent monitoring system has been proposed.Non-overl.apping multi-camera surveillance system is the important research content in wide area surveillance system.Person re-identification as the fundamental issue and main problem in non-overlapping multi-camera surveillance system has become a hot research.Person re-identification is to identify whether a given person in the surveillance network has emerged in other cameras.This is a new and interdisciplinary field of study,there is widespread application prospect.Due to lack of time limits and path information,person re-identification mainly relies on extracting appearance characteristics of target and building similarity measure criteria between features.Because of the color characteristics having spatio-temporal continuity and strong robustness to attitude change is often used as the main feature combined with other features to describe pedestrian targets.But the color is easily influenced by many factors such as camer.a sensor characteristics,illuminant and ambient light.That makes the same target between different cameras have great color difference.Therefore it is extremely important and necessary to make color consistency between cameras.That is to find a characterization which is stable,robustness to camera and illumination change and can reflect the intrinsic color of target.Similar targets exist;occlusion between the objects and background adhesions will also affect recognition result.This paper deeply studies these issues.The main research results as follows:Firstly,color consistency between multiple cameras is realized.This paper studies two color consistency algorithms based on color appearance and physical model of imaging.On the one hand,this paper coImbine the chromatic adaptation theory,CIECAM02 color appearance model and iCAM image color appearance model,transform pedestrian under different circumstances to standard lighting conditions and get target true color to eliminate the impact of light and ambient light.On the other hand,the reflective properties of the target,scene illumination and camera response are separ.ated by imaging model.Then the reflection characteristic image of the target can be obtained.The influence of the camera performance to color can be eliminated by solving the camera response inverse function to recover the image irradiation.Then according to the intrinsic image decomposition theory,the reflectance component and shading component are separated,and the reflection image that represents the target surface characteristics can be obtained.The availability of color consistency algorithms has been verified by experiments.This can provide support for person re-identification.Secondly,according to practical issues exist in person re-identification such as attitude change,perspective change,and occlusion and so on,three algorithms have been proposed in this paper.The first algorithm is based on color constancy theory.Appearance characteristics for obj ect recognition are established by color invariants which has nothing to do with illumination.Color invariants consist of edge color invariant based on Gaussian spectrum model and moment invariants based chrominance space.The second algorithm uses the CSIFT feature.It combines the hue invariant and scale invariant features SIFT.Experiment results show that the two algorithms are feasible,and can resist the change of lighting conditions,perspective and attitude.In order to oppose the effects of occlusion and background adhesion,the third algorithm is proposed.The human target is divided into three parts;include head,torso and leg.The overall color feature,local shape and local color feature of the target main color and the texture feature are extracted.Then the person re-identification based on multi-feature fusion is realized.Experiments in VIPeR database and ETHZ database confirm the algorithm is effective.At the same time it has been confirmed by VIPeR database,that the recognition rate can be improved by color consistency between different cameras.Finally,in order to solve the problem that the visual similarity target seriously impact the target recognition rate,the person re-identification algorithm based on uniform color space and salience region is proposed by analyzing the visual similarity objects.Based on the visual consistency of different uniform color space by visual experiments,the IPT color space having the best results is selected.In the color space the salient features of the target is detected based on Bayesian theory and statistical characteristics of the local region,and then the salience features are extracted.On this basis,the person re-identification is realized by combining the features based on color spatial distribution and salient region of target.Experimental results show that the algorithm used to visual similarity pedestrian recognition is feasible and effective.