Person Re-Identification Based On Color Names Feature Augmentation

With the development of society,more and more cameras have been installed in public places such as shopping malls,subways,campuses,ports and airports,forming a large video surveillance network.Person re-identification is a technique for judging whether person under the non-overlapping cameras are the same person.Due to different parameters,different installation angles and different surrounding environments of different cameras,person under different cameras may be affected by factors such as viewing angle,illumination and posture,and the same person may not be similar and the different person may be similar.In order to solve the problems that may occur in the person reidentification,this paper studies from two aspects: feature extraction and metric learning.The proposed method first converts the three color channels in the color features commonly to the eleven color channels which are closer to the real world color.Each color channel corresponds to a color name,and calculates the probability that each color name appears at each pixel position.In addition,in this paper,the color name with the highest probability of appearing in eleven colors of each pixel position directly replaces the pixel of that position,so as to solve the influence of illumination and surrounding environment on the color names feature,and make the color name feature more robust to illumination and environment.The quadratic discriminative analysis metric method learns a projection matrix common to all cameras,so its generalization ability is strong,but it ignores the inherent feature distortion caused by the inherent properties of each camera.These distorted features may have discriminatory information,so the discriminative ability of the method is weak.In this paper,the feature augmentation method is used to improve the quadratic discriminative analysis metric.While ensuring its generalization ability,consider the feature distortion caused by the inherent properties of each camera to improve its final discriminating ability.Finally,two datasets,VIPeR and CUHK01,which are widely used in person re-identification,are validated.The experimental results show that the feature extraction method and the improved metric learning method proposed in this paper can improve the recognition rate of person re-identification.