Robust Stereo Road Image Segmentation Using Threshold Selection Optimization Method Based on Persistent Homology

Abstract

This paper introduces a novel method for road target segmentation in the context of unmanned driving based on stereo disparity maps. The proposed method utilizes topological persistence threshold analysis to address the challenges associated with selecting appropriate threshold values. The approach involves converting 3D vehicle road images into uv-disparity maps, extracting road planes using v-disparity maps, and calculating occupancy grid maps using u-disparity maps. Persistence diagrams are then constructed by generating segmentation results under various threshold parameters. By establishing persistence boundaries in these diagrams, the most significant regions are identified, enabling the determination of robust segmentation thresholds. Experimental validation using KITTI stereo image datasets demonstrates the effectiveness of the proposed method, with low error rates and superior performance compared to other thresholding segmentation methods. The research holds potential for application in autonomous driving systems.