Deep feature representation and multiple metric ensembles for person re-identification in security surveillance system

Abstract

With the increasing concern about social public security and the development of large scale data storage technology, person re-identification in security surveillance system becomes a hot topic. Large variations in viewpoint and lighting across different camera views could change the appearance of the person a lot, which makes person re-identification still a challenging problem. Therefore, developing robust feature descriptors and designing discriminative distance metrics to measure the similarity between pedestrian images are two key aspects in person re-identification. In this paper, we propose a method using both deep learning and multiple metric ensembles to improve the performance of the re-identification. Firstly, we jointly use the various datasets to train a general Convolutional Neural Network (CNN) which is employed to extract the deep features of training and testing set afterwards. The deep architecture makes it possible to learn more abstract and internal features which are robust against the variations in viewpoint and lighting. Then we utilize the deep features of the training set to learn the specific distance metric of different datasets and combine it with Cosine distance metric together, multiple metric ensembles can measure the similarity between different images in a more comprehensive way. Finally, extensive experiments demonstrate that our method can improve the recognition performance effectively when compared with the state-of-the-art methods.