A YOLOv8 Segmentation Approach for Detecting Unstructured Road Boundaries in Rural Terrains Using Dash-Camera

Unmarked road boundaries have always been a significant challenge for autonomous and assistive driving systems, especially in rural areas where standard lane markings are absent. This paper presents a segmentation-based approach for detecting unmarked road boundaries using a YOLOv8 model trained on custom dash-camera footage captured in varied daytime lighting conditions. The dataset is manually annotated with polygon-based segmentation using the computer vision annotation tool to capture the diverse appearances of rural road edges, including irregular terrains and undefined boundaries. These annotations were initially exported in common-objects-in-context segmentation format and later converted to the YOLOv8 segmentation format. A medium-sized YOLOv8 segmentation model was trained locally on this dataset. The objective is to build a system that could work in real-time and accurately present the unmarked road boundaries in unpredictable rural terrains. The model showed promising results on the validation set, performing consistently even in unpredictable terrains and uneven road surfaces, and a real-time prototype was also developed to test the model live, using phone-camera video as input for inference and visualization. This approach demonstrates the potential for a lightweight deep learning model to support real-time road boundary detection in unstructured, low-infrastructure driving environments.