Research On Key Technology Of Person Re-identification

Person re-identification?RE-ID?is a challenging problem focusing on pedestrian matching and ranking across non-overlapping camera views.It remains an open problem although it has received considerable exploration recently,in consideration of its potential significance in security applications,especially in the case of video surveillance.Meanwhile,the resolution of image obtained by the surveillance camera is low,view perspective and the pedestrian posture are easy to be changed,and the background is complex,so that the same pedestrian has different appearances in different surveillance cameras.In order to solve these problems effectively,this paper combines the feature extraction and feature selection methods to research RE-ID from the feature representation.The main research contents include the following aspects:?1?Stripe and block features fusion.Firstly,the stripe feature is extracted by HSV,LAB,RGB and YCrCb four color characteristics and Gabor filter,and GOG descriptor is used to extract block feature,then those features are concatenated into a whole feature vector,which is projected into discriminative null space to reduce their feature dimension.Finally,Euclidean distance is applied to compute personal distance to achieve re-identify person.Experimental results show that,rank1 recognition rate of the proposed method could achieve 52.7%,72.2% and 59.7% on VIPeR?Prids₄50s and CUHK01 database respectively,which shows that the proposed method can sufficiently describe personal picture feature,and has a strong robustness for environment which can effectively improve the recognition rate.?2?Multi-color features with D-optimal partial least squares feature selection.Firstly,multi-color features,which consist of HSV,LAB,RGB and nRnG color features,were extracted and concatenated into a whole feature vector.Then D-optimal Partial Least Squares features selection was adopted to select an optimal feature subset that could minimize the variance of the regression model.Finally,an asymmetric distance model for similarity matching was utilized to learn discriminative feature from different perspective.Experimental results show that the rank1 performance of the proposed method were 46.27%,61.87% and 64.47% respectively on the VIPeR,Prid₄50s and CUHK01 databases,which has achieved the state-of-art performance.?3?Ordinal locality feature selection with divide and fusion.The dimensionality of extracted features is more than thousands,so that selecting effective pedestrian information with unsupervised ordinal locality suffers from the curse of dimensionality.To solve this problem,firstly,the extracted multiple high-dimensional features were divided into several sub-features respectively.Then each sub-feature was applied to obtain better feature and reduce dimensionality with ordinal locality feature selection.Finally,the sub-features of each feature were cascaded respectively with serial feature fusion,and the multiple cascaded features were fused by weighted into one vector for re-rank with parallel feature fusion.Experimental results on four challenging person reidentification datasets demonstrate the effectiveness of proposed method.Especially,the proposed method outperforms the state-of-the-art methods,and reduces computational complexity.