Non-linear metric learning with multiple features for person re-identification

In video surveillance, person re-identification across disjoint camera views has important applications. Many factors make it difficult to tackle such as the varieties of varying lighting conditions, viewing angles, body gestures, background clutter and occlusion. To address the problem, several methods exploited combination or the fusion of multiple features, or built models. However, in these methods, contributions of different features were determined too subjectively or the use of differences between different samples is inadequate in building model. For these problems, a novel method is proposed by us, which combines multiple kernel learning (MKL) and distance metric learning (DML) to fully employ the information of several different features in the process of re-identification and build the most descriptive and robust model. Furthermore, the distance metric learning method is improved according to the practical and the related parameters of kernels can be self-selected by our method. Experiments are conducted on benchmarking dataset, and the experimental results suggest that our approach achieves encouraging performance.