Affine Transform for Skew Correction Based on Generative Adversarial Network Method for Multi-Camera Person Re-Identification

Abstract

In intelligent video surveillance system, person re-identification is a key technology. In order to address the problem, the decrease in performance of person Re-Id lead by the skew pedestrian images, this paper proposes the affine transform for skew correction based on generative adversarial network (GAN) method for multi-camera person re-identification (Re-Id). Firstly, an effective GAN is proposed to guide the spatial transformer network (STN) to learn affine transform parameters for skew correction in an adversarial way, and STN is adopted as the preprocessing model for Re-Id to reduce influence of variations in person posture. Then, features are extracted by a deep convolutional neural network from input images which are corrected by STN, and finally results can be obtained by measuring similarity between features. Besides, in the proposed GAN, a classification model and related loss functions are introduced to reduce the damage to the key features of pedestrian during skew correction. The effectiveness of the proposed method is verified by experiments conducted on the skew pedestrian dataset.