In response to the challenges posed by multi-scale object recognition, limited shape adaptability, and insufficient contextual awareness in road obstacle detection under complex urban environments, an upgraded detection model named ERLM-YOLOv8 is developed in this work, leveraging the YOLOv8 framework as its foundation. An Enhanced Adaptive Spatial Feature Fusion (EASFF) head is integrated into the model to enable dynamic fusion of hierarchical features, effectively improving multi-scale target recognition. The RO-DSC module is incorporated into the architecture to strengthen the modeling of irregular obstacle shapes, improving adaptability to diverse object shapes. Additionally, the backbone network is augmented with the SPPF-LSKA attention mechanism to capture contextual dependencies across global scales, while we employ the MPDIoU loss function to enhance the bounding box regression procedure. Experiments conducted on a custom dataset containing 10,686 road images with 17 obstacle categories demonstrate that the proposed model obtains 86.8%, surpassing the baseline YOLOv8n (81.6%) by 5.2 percentage points. Comparative results further show that ERLM-YOLOv8 outperforms YOLOv10 and YOLOv11 by 8.1% and 6.0%, respectively, while maintaining a fast inference speed of 11.7 ms per frame. It is highlighted by these results that the model excels at reducing both missed and false detections of multi-scale obstacles, presenting an accurate and low-latency approach for real-time obstacle perception in autonomous driving frameworks.