Person Re-identification Based On Metric Learning And Deep Learning

Due to the rapid development of video surveillance network, it is a great challenge to process the huge video data with traditional artificial video analysis method. The intelligent video analysis method based on person re-identification has become the focus and hot spot in the field of computer vision. Person Re-Identification is a method which verify the pedestrian captured by the camera at different location and times is the same person. At present, the methods of person re-identification can be divided into two categories: one is based on feature representation, the other is based on model learning. By designing the feature descriptors, feature representation methods extract the distinguishing features of pedestrian, Instead of automatically learning features from the data. On model learning method, learning feature transforms method is complex and has a great influence on the illumination and camera parameters, Distance metric learning of the complex distribution is not strong and robustness. By combining the image color space characteristics with the adaptive expansion and contraction of the distance metric learning, a new person re-identification method based on adaptive distance metric learning method is proposed, which is robust to the distribution of complex heterogeneous data. Furthermore, a deep learning method is constructed to tackle the problem of adaptive feature extracting, and use combine classification to research the person re-identification. The specific research is described as follows:(1) Based on the characteristics of the image color space, the HSV sub features of the pedestrian image are extracted by the blocks, and then the feature vectors are constructed to characterize the pedestrian image. This paper introduces an adaptive shrinkage and expansion distance metric learning method, which combines the two methods to form a person re-identification system.HSV color space is more in line with human color perception, HSV color feature distribution does not change with the deformation, rotation, translation in the images of. In this paper, the HSV features and the improved M-SEAML(Manhattan Distance-Shrinkage Expansion Adaptive Metric Learning) algorithm is combined to construct HSV feature and adaptive shrinkage expansion distance metric learning person re-identification system. Experimental result in VIPeR[3] database shows that the Rank1 index of the proposed method is equal to 19.9367%.(2) By fine-tuning training the deep convolutional networks model, the person re-identification method based on the deep learning is proposed to use the deep learning model for feature extraction.In order to tackle the problem of finite amount of train data, Caffe based model is used, which is robustness with illumination, rotation changes. After feature extraction, the person re-identification is transformed the problem as a binary Softmax classification with pairwise constraints. Experimental result in VIPeR[3] database shows that the Rank1 index of the proposed method is equal to 30.4589%.