Person Re-identification Algorithm Combining Spatio-temporal Apparent Feature Fusion With Feature Matching

Person Re-identification(Person Re-ID)is a key technology of intelligent video surveillance system,which uses visual and motion information to achieve pedestrian target matching in non-overlapping areas,extending single-camera to multi-camera network.While multi-camera network provides rich monitoring information,there are also vision differences between cameras.How to extract effective features in complex video data and match the feature discriminatively are the main directions of the research.This paper explores the person re-id further from two perspectives,one is feature extraction of video pedestrian targets and the other is feature matching.Firstly,aiming at the problem that apparent features can't provide movement information and limit the expressive power,this paper proposes a person re-identification model based on spatio-temporal appearance feature and Cross-view Quadratic Discriminant Analysis(XQDA).The first step is to extract spatio-temporal gradient direction histogram HOG3 D feature of pedestrian,and then construct HOG3 D with color histogram,texture and other appearance features.The next step is to realize metric learning by using the Crossview Quadratic Discriminant Analysis algorithm which is improved based on the classical distance measurement algorithm(Keep It Simple and Straight Metric Learning,KISSME).At last,the feature distance measurement and sorting are carried out to complete the reidentification process.Secondly,in order to solve the problem of large differences between characterized views caused by different camera parameters,a camera correlation aware feature augmentation based on Marginal Fisher Analysis is proposed.At the beginning,on the basis of extracting pedestrian's spatio-temporal appearance feature,the pedestrian data collected by different cameras are used to learn a camera correlation function.Then,on the basis of the obtained correlation,the view-general features are augmenteded adaptively with MFA algorithm in the new adaptive space.At last,measure the Euclidean distance between features and sort them.Finally,manual features are difficult to represent pedestrian abstract features,which can't deal with dramatic posture changes well.In order to solve the problem,this paper proposes a CNN feature extraction method.At the beginning,response intensity histogram and ordinal histogram are extracted based on the convolution layers' s feature maps of AlexNet,a convolution neural network already trained in image dataset ILSVRC 2012,and then the two histograms are fusioned to form a reliable depth feature descriptor.On the basis,the validity and generalization performance of CNN feature are verified by combining camera-related feature enhancement algorithm.