Dynamic Matching of Local Features for Re-Identification of Pedestrians

Abstract

This paper presents a person re-identification (reid) method based on a convolutional neural network (CNN). There are many CNNs for the re-id problems, where few of them considered the image properties in real-world situations that the images could be deformed due to the perspective view of surveillance cameras and that pedestrian detectors do not give perfect bounding box. In this paper, we address the problem of perspective view and incomplete bounding box by proposing a new network architecture and metric learning method. Specifically, we compensate for the vertical and horizontal misalignment due to the incomplete bounding boxes of the pedestrian detector, and also horizontal squeezing that was not considered in the existing algorithms. For this, we partition the bounding box of pedestrian detection results into M horizontal region and N vertical regions. Then, we apply a dynamic matching technique in both horizontal and vertical directions to compensate for the effects of unfit bounding boxes and squeezed appearance due to the perspective views. The partitioning and dynamically matched distance are also considered for the metric training of CNN. We compare our method with state-of-the-art ones and validate the improved performance.