TLDR: SBP-YOLO is a new lightweight, real-time AI model built on YOLOv11 for detecting speed bumps and potholes. It uses efficient components like GhostConv, VoVGSCSPC, and LEDH, along with a hybrid training strategy, to achieve high accuracy (87.0% mAP) and fast inference (139.5 FPS on Jetson AGX Xavier), making it ideal for intelligent vehicle suspension systems.
Modern vehicles, especially new energy vehicles, are increasingly focusing on passenger comfort. A key aspect of achieving this comfort is the ability to accurately and quickly detect road irregularities like speed bumps and potholes. This real-time information is crucial for advanced suspension systems that can adjust proactively to road conditions.
Traditional methods for detecting these road anomalies often rely on vehicle dynamics or expensive sensor setups, which can be computationally intensive and lack the predictive capability needed for intelligent suspension control. However, recent advancements in deep learning, particularly with computer vision techniques, offer a more proactive and accurate solution.
A new research paper introduces SBP-YOLO, a lightweight and efficient object detection framework designed specifically for real-time detection of speed bumps and potholes. Built upon the YOLOv11 architecture, SBP-YOLO is optimized for deployment on embedded systems, which are common in vehicles due to their limited computing power.
The SBP-YOLO model incorporates several innovative components to achieve its impressive performance. It uses GhostConv modules for efficient computation, reducing the processing and memory costs without sacrificing the ability to extract important features. To enhance the model’s ability to understand features at different scales, especially for small and distant objects, it integrates the VoVGSCSPC module. Furthermore, a Lightweight and Efficiency Detection Head (LEDH) is proposed to handle early-stage feature processing efficiently, ensuring accuracy for small targets while keeping computational overhead low.
To make the model robust and accurate, particularly for small and distant targets that are often challenging to detect, a unique hybrid training strategy was employed. This strategy combines NWD loss for precise localization, knowledge distillation to transfer learning from a larger model, and Albumentations-based weather augmentation to simulate various challenging conditions like motion blur, lighting changes, rain, and snow. This comprehensive approach significantly improves the model’s ability to generalize to real-world scenarios.
Also Read:
- Enhancing Pedestrian Detection Accuracy with a Comprehensive Proposal Refinement Algorithm
- New Deep Learning Models Enhance Safety at Highway-Railway Crossings
Experimental results demonstrate the effectiveness of SBP-YOLO. It achieved an 87.0% mAP (mean Average Precision), which is a significant improvement over the baseline YOLOv11n model. More impressively, when optimized with TensorRT FP16 quantization, SBP-YOLO achieved a remarkable 139.5 frames per second (FPS) on the Jetson AGX Xavier platform. This high inference speed and computational efficiency make it highly suitable for real-time road condition perception in intelligent suspension systems, ultimately leading to a smoother and safer driving experience. You can read the full research paper here.
