Anti-occluded Person Re-identification via Pose Restoration and Dual Channel Feature Distance Measurement

In real scenes, persons are often blocked by obstacles. The purpose of occluded person re-identification is to identify the occluded persons in the non-shared view camera. In this paper, we propose a new framework, anti-occluded person re-identification model via pose restoration and dual channel feature extraction (PRAO). The network is divided into two modules: person pose repair module (PPR) and dual channel feature extraction module (DCFE). (1) In the person pose repair module, the instance segmentation network is used to detect the occlusion in the person image, and then the pre-trained edge smoothing GAN (e-GAN) is used to repair the person image. (2) In the dual channel feature extraction module, we change the original single channel person prediction structure into a dual channel person matching structure, which can accurately align the person features of the two channels, reduce the noise generated by image generation, and improve the identification accuracy of persons in the case of occlusion. Finally, a large number of experiments on occluded and non-occluded datasets show the performance of the method.