Person re-identification is a challenging task aiming to identify the same person across different cameras. However, most of existing image-based person re-identification methods neglect the spatial and temporal constraint, the information in neighbor frames of each person image is rarely exploited by previous studies. In this paper, we propose a novel neighbor frames mining framework (NFM) to exploit the spatial-temporal information. For each gallery image, we use a dynamic programming-based global optimal tracking method to search images of the same person in its neighbor frames. From those images, the image features extracted by the shared convolutional neural network (CNN) in the constructed neighbor sequence are merged via an attention weighted averaging technology. To this end, a novel supervised attention mechanism is designed for dealing with tracking errors. The final feature with multi-view and robust information is used for matching. Experimental results show the superiority and efficiency of the proposed method on two benchmark datasets including DukeMTMC-reID and PRW.