Reliable obstacle detection in complex urban environments remains a fundamental challenge for autonomous driving, as monocular object detectors often lack robust depth perception — leading to inaccurate localization and high false-positive rates. To address this limitation, we propose BIF-YOLOv8, a real-time obstacle detection framework that effectively integrates binocular geometric reasoning with the semantic strength of an enhanced YOLOv8 architecture. Our approach introduces a Binocular Information Fusion (BIF) that leverages stereo disparity to perform depth-aware confidence reweighting and enforce geometric plausibility on detection outputs. In parallel, we adopt the Weighted Intersection over Union (WIoU) loss to improve bounding box regression by adaptively focusing on hard-to-localize samples during training. Evaluated on the KITTI benchmark, BIF-YOLOv8 achieves state-of-the-art precision of 97.55%, significantly suppressing false alarms while maintaining competitive recall. The full model attains an mailto:mAP@0.5 of 89.75%, outperforming the baseline YOLOv8 by 7.41 percentage points. Moreover, it operates at 32.8 frames per second (FPS) with 7% fewer FLOPs and 10% fewer parameters, demonstrating a favorable trade-off between accuracy, robustness, and efficiency. These results confirm that BIF-YOLOv8 provides a practical, lightweight, and reliable solution for safetycritical obstacle perception in intelligent transportation systems.
Wang et al. (Fri,) studied this question.