Research Of Objects Association In Non-overlapping Multi-camera

With the popularity of wide-area video surveillance, the research of multi-camera intelligent video surveillance has become a hot topic in computer visual. Given the financial cost and civil privacy concerns, it is impractical to equip an area with an entire camera network that has no blind spots. This makes it difficult to track an individual across different camera views. Objects association in non-overlapping multi-camera, a.k.a. person re-identification, is mainly used to remove impacts on object tracking, which are caused by visual differences between cameras and time-space uncertainty of blind spots. In non-overlapping multi-camera surveillance system, objects association is a key issue for object tracking. In this paper, we make an in-depth study of non-overlapping multi-camera objects association, and the main work and innovations are as follows:(1) In this paper, we propose a re-identification method that uses Nonlinear Ranking with Feature Difference (NRFD). Instead of trying to eliminate the effects of visual differences between cameras as many re-identification methods done, we make full use of targets’differences to build a binary classifier. We then achieve re-identification by ranking target candidates through a support vector machine with a nonlinear kernel based on radial basis function. Furthermore, we propose to pre-cluster the training images using the affinity propagation clustering algorithm, and select representative images to form negative training instances. This avoids an imbalanced training problem caused by large negative training instances number and small positive training instances number. Extensive experiments are conducted on three public benchmark datasets, and the results demonstrate the state-of-the-art performance of the proposed method.(2) Considering the JENSEN-SHANNON (JS) kernel based on information measure can handle color histograms shifted by illumination changes, a JS kernel discriminant analysis (JSKDA) model is formulated to solve objects association in this paper. We first map each image into a high-dimensional space constructed by JS kernel, and fully extract the identifiable information. Then we apply the improved local Fisher discriminant analysis to reduce the impact of outliers, and complete the objects association with lower error rate.(3) We introduce D-S evidence theory to merge time-space information, NRFD model, and JSKDA model. In this way, we avoid the conflicts among the models and increase the precision of objects association.