Research Of Problems Of Non-overlapping Multi-Camera Pedestrian Re-identification Based On Multiple Kernel Learning

Recently, with the development of intelligent video surveillance technology, the problem of person re-identification across disjoint camera views has become one of the hottest researches in the field. The main task of this problem is to re-identify a specific individual who has appeared in one camera when he or she is captured by another camera without overlapping views. The issue has the following several difficulties: varieties of varying lighting conditions, lack of spatial and temporal constraints and uncertainty of people gestures. To solve the problem, a novel approach of modeling based on multiple features of person appearance is proposed in this paper. And main work and contributions of this paper focus on the following aspects:1. In the respect of cameras relationship modeling, a distance metric learning approach called Relaxed Margin Components Analysis (RMCA) is proposed. Considering data distribution in one-shot and multiple-shots dataset, the "margin" function of Large Margin Nearest Neighbors (LMNN) is redefined to make it relaxable in terms with data distribution. This method is able to achieve good re-identification performance and convergences quickly.2. A Kernel RMCA (KRMCA) approach is proposed. Considering that samples are more separable in high-dimensional space, the pedestrian samples are projected into it, and then are used to model through RMCA algorithm. Some common kernel functions are analyzed. And a conclusion is drawn, that the Jensen-Shannon kernel is the most suitable one for specific person features. Compared to RMCA method, experimental results show that KRMCA further enhances the accuracy of re-identification.3. A method of person re-identification combined with multiple features fusion is proposed. To achieve the importance automatically selection of different features, combining the idea of Multiple Kernel Learning (MKL), KRMCA is improved to Multiple Kernels RMCA (MKRMCA). Based on it, a novel selection approach of kernel width parameter is proposed and the kernel width parameter could be generated adaptively. The results of experiments show that this method achieves the higher performance than KRMCA, especially the first few ranks of CMC.