Improvement of panoptic segmentation method for urban road

Abstract

When learning and studying the panoptic segmentation method upsnet, in order to better apply it in the intelligent driving scene, the following improvements are made to the algorithm: 1. Aiming at the problem that the scale difference of different types of targets in the traffic scene is too large, the feature extraction network of the network is improved, and the upsnet panoptic segmentation network combined with recursive feature pyramid is proposed. 2. Aiming at the occlusion problem between different categories in panoptic segmentation task, an occlusion processing model is added to upsnet to solve the occlusion problem. The improved algorithm is compared with upsnet and other excellent panoptic segmentation networks on cityscapes data set and the panoptic segmentation data set labeled in this paper. The experimental results show that the evaluation index PQ (panoptic quality) has been greatly improved, and the improved network is more suitable for intelligent driving scenes.