Road Detection without Lane Markings for Autonomous Vehicles

This paper presents a simplified training strategy for drivable area detection, focusing on road environments where lane markings are missing or unclear. Based on the YOLOParchitecture, the proposed method removes the lane detection branch and adopts a two-stage training process. The model is first trained on object detection to allow the encoder to learn useful visual features. In the second stage, segmentation is introduced, and both tasks are trained together in a multitask setup. This helps the model maintain detection performance while learning to identify drivable areas accurately. This design reduces task interference and helps the model better adapt to unstructured roads. The approach aims to improve model accuracy and robustness while maintaining architectural efficiency for real-world applications. Evaluation on a custom dataset featuring unstructured road conditions demonstrated that the proposed model achieved notable segmentation accuracy (92.60% mIoU) and reliable object detection (91.30% recall, 68.10% mAP@0.5).