Research On Person Re-identification Based On Multi-features

Person re-identification is a hot spots and also an open problem in computer vision.It has a wide prospect in many application domains,such as human–robot interaction scenario,security and surveillance and forensics systems.Study on the feature representation and the metric learning of person re-identification has very important practical and theoretical significance.In order to improve the accuracy of person re-identification,this thesis focuses mainly on the feature representation and the similarity between the features.The main contributions are as follows:1)The current methods for person re-identification either use feature of the whole person or the features of the body-parts.In order to make full use of the feature of the whole person and the features of the body-parts,we propose a method of Multi Parts-based Metric Learning.The method learns a distance metric for both the feature of the whole person and the features of the body-parts,then the similarity learned by the metric will be combined with different weight to get the final similarity between the person.Compared to the methods with feature of the whole person or the features of the body-parts,the Multi Parts-based Metric Learning method improves the accuracy for person re-identification by 3%-9% on datasets PRID2011 and VIPe R on average.2)The features used for person re-identification either the color,texture and interest point or the fusion of them,they are low-level features.The high-level attributes are proved to be efficient,but the attributes are hard to get.Color Name is a mid-level feature which is easy to get.To make advantage of the low-level and mid-level features,the low-level and mid-level features are fused with different weight by the discriminate ability for person reidentification.The experimental results show that the fusion of low-level with mid-level semantic feature improves the accuracy on datasets PRID2011 and VIPe R on average...