This paper proposes a SBP-YOLO framework for efficient detection of speed bumps and potholes in embedded systems. Based on YOLOv11n, we integrate GhostConv and VoVGSCSPC modules to reduce computational complexity and enhance multi-scale semantic features. We improve small object detection through P2-level branching and minimize overhead while maintaining accuracy using a Lightweight Detection Head (LEDH). A hybrid training strategy combining NWD loss, BCKD knowledge distillation, and Albumentations-based augmentation enhances robustness under diverse road and environmental conditions. Experimental results show that SBP-YOLO achieves 87.0% mAP, a 5.8% improvement over the YOLOv11n baseline. After TensorRT FP16 quantization, the implementation runs at 139.5 FPS on a Jetson AGX Xavier, demonstrating a 12.4% speedup over the P2-enhanced YOLOv11.
