Scholarlink

<p>Preview-based vehicle suspension systems enhance ride comfort and handling stability through proactive road anomaly detection and suspension control. To ensure adequate time for system response, higher vehicle speeds require detecting obstacles at greater distances, where they appear as smaller objects in the image space. This creates a critical challenge for embedded platforms: achieving high detection accuracy for distant small targets while maintaining real-time performance under computational constraints. To address this, we propose SBP-YOLO, a lightweight YOLOv11-based detector that integrates GhostConv and VoVGSCSPC for efficient multi-scale feature extraction. The model incorporates a P2 branch for small-object sensitivity and a Lightweight and Efficient Detection Head (LEDH) that reduces the computational overhead introduced by the P2 layer while maintaining accuracy. A hybrid training strategy employing normalized Wasserstein distance (NWD) loss, knowledge distillation, and albumentations augmentation enhances robustness to environmental variations. Experimental results demonstrate that SBP-YOLO achieves 87.0% mAP, exceeding YOLOv11n by 5.8%, with notable improvements in distant small-object detection. After TensorRT FP16 quantization, the model achieves 139.5 FPS on an NVIDIA Jetson AGX Xavier, representing a 12.5% speed improvement compared to the P2-enhanced YOLOv11, while simultaneously reducing energy consumption by 14.8%–174.84 mJ per frame. These findings validate SBP-YOLO as an efficient real-time solution for intelligent suspension systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SBP-YOLO: a lightweight real-time model for detecting speed bumps and potholes toward intelligent vehicle suspension systems

  • 期刊论文

  • 发表日期：

    2026年1月18日

查看全文
  • Chuanqi Liang,
  • Jie Yuan,
  • Linlin Gou,
  • Lei Luo,
  • Jie Fu,
  • Miao Yu

摘要

Preview-based vehicle suspension systems enhance ride comfort and handling stability through proactive road anomaly detection and suspension control. To ensure adequate time for system response, higher vehicle speeds require detecting obstacles at greater distances, where they appear as smaller objects in the image space. This creates a critical challenge for embedded platforms: achieving high detection accuracy for distant small targets while maintaining real-time performance under computational constraints. To address this, we propose SBP-YOLO, a lightweight YOLOv11-based detector that integrates GhostConv and VoVGSCSPC for efficient multi-scale feature extraction. The model incorporates a P2 branch for small-object sensitivity and a Lightweight and Efficient Detection Head (LEDH) that reduces the computational overhead introduced by the P2 layer while maintaining accuracy. A hybrid training strategy employing normalized Wasserstein distance (NWD) loss, knowledge distillation, and albumentations augmentation enhances robustness to environmental variations. Experimental results demonstrate that SBP-YOLO achieves 87.0% mAP, exceeding YOLOv11n by 5.8%, with notable improvements in distant small-object detection. After TensorRT FP16 quantization, the model achieves 139.5 FPS on an NVIDIA Jetson AGX Xavier, representing a 12.5% speed improvement compared to the P2-enhanced YOLOv11, while simultaneously reducing energy consumption by 14.8%–174.84 mJ per frame. These findings validate SBP-YOLO as an efficient real-time solution for intelligent suspension systems.