Research On Key Technology Of Person Re-identification In Non-overlapping Surveillance Scenarios

Matching persons observed from non-overlapping camera views, known as person re-identification(re-id), aims to identify the given target person image taken from non-overlapping surveillance system. This technique has become one of the most important tasks in video surveillance systems, such as human retrieval and activity analysis.Nevertheless, this is a very challenging task. The same person observed in different camera views undergoes significant variations in viewpoints, poses, illumination, background, camera settings, which usually make intra-personal variations even larger than inter-personal variations. Under such circumstances, how to establish an efficient and robust person re-id method causing extensive attentions of researchers in recent years.In this paper, the problem of person re-identification across non-overlapping camera views is discussed. The main achievements are as follows:1) Aiming for objects' unsteadied inter-salience properties in person re-identification, a new algorithm of person re-identification is proposed on the base of Multi-Salience Fusion.Firstly, two-stage Manifold Ranking(MR) is established to obtain image intra-salience. Then a more accurate Multi-Salience Fusion description method is presented after combining intra-salience with existing inter-salience. Therefore, a more accurate salience description for objects is formed by combining it with inter-salience. Compared with the similar algorithms,the method can describe the salience of objects more accurately, reaching high re-identification rate.2) Aiming for the problem of inconsistent salience properties between matched patches in person re-identification, a multi-directional salience similarity evaluation approach for person re-identification based on metric learning is proposed on the basis of multi-salience fusion. Firstly a distribution analysis for four kinds of saliency of matched patches is taken,and the visual similarity between matched patches is established by multi-directional weighted fusion of salience. Then the weight of saliency in each direction is obtained using metric learning in the base of Structural SVM Ranking. Last the visual similarity between image pairs is established by multi-directional weighted fusion of salience. The approach considers not only the salience of matched patches in each direction, but also the weight of directional saliency. It improves the discriminative and accuracy performance of re-identification. Compared with the similar algorithms, the method achieves higher re-identification rate with more comprehensive similarity measure.3) Aiming for metric learning based person re-identification using principal component analysis, which easily causes loss of classification information, a new person re-identification algorithm is proposed based on incremental linear discriminant analysis. Firstly by using LDA(Linear Discriminant Analysis) method, samples can be projected into low-dimension subspace on the basis of maintaining the maximize classification information. And a distance metric is learned in low-dimension subspace as well as the projective transformation matrix.Then Mahalanobis matrix of samples' low-dimension subspace is obtained by transformation matrix and the covariance matrix of the original sample space. Last it introduces metric learning method to enable the metric learning model updating according to the training set ofthe new training samples. The proposed method not only considers retaining sample classification information in projection subspace, but also considers the update ability of the metric model, which can enhance the accuracy and extensibility of the algorithm. Compared with the similar algorithms, the method can reach a high person re-identification rate.