자율주행 트램을 위한 중심선 기반 Ego-Path 후보 분할 기법

초록

This paper proposes a method for reliably detecting and instance-wise segmenting multiple ego-path candidates in straight and branching sections from camera images. The proposed method enhances the visual separation between adjacent tracks through perspective transformation-based preprocessing and introduces a centerline heatmap branch into an instance segmentation model to jointly learn track regions and structural characteristics, including continuity, directionality, and branching structures. Experimental evaluations using benchmark datasets demonstrate that perspective transformation and centerline-based structural learning contribute to improving the segmentation performance of ego-path candidates in complex branching sections. Furthermore, experiments conducted on a real tram test track under normal, rainy, and snowy conditions confirm that the proposed method can effectively detect drivable track candidates under diverse environmental conditions.

키워드