Enhancing Real-Time Road Object Detection: The RD-YOLO Algorithm With Higher Precision and Efficiency

Abstract

To address the challenges of low detection accuracy, high miss rates, and inefficient model computation in complex road scenes, this study introduces RD-YOLO, an enhanced version of the YOLOX algorithm. First, we replace certain traditional two-dimensional convolution (Conv2d) modules in the backbone and neck networks with Ghost Convolution (GhostConv) modules and incorporate a Multidimensional Collaborative Attention (MCA) mechanism. This integration improves the model’s feature extraction capabilities while reducing the number of parameters. Second, we design a high-resolution detection branch to enhance the accuracy of small-scale object detection. Finally, we introduce MPDIoU as the new bounding box regression loss function, replacing the original loss function to accelerate algorithm convergence and improve target localization accuracy. Experimental results on the BDD100K dataset demonstrate that RD-YOLO increases the mean Average Precision (mAP) by 2.1% compared to the baseline model, while also reducing the number of parameters and FLOPs by 13.6M and 38.2G, respectively. Compared to other mainstream object detection methods, RD-YOLO achieves a mAP of 62.8%, with a parameter count of 35.6M and a Frames Per Second (FPS) of 106.7, showcasing significant advantages and better alignment with the demands of road scene object detection. Moreover, on embedded devices, RD-YOLO sustains a mAP of 62.8%. Although the FPS drops to 35.6, it still outperforms YOLOX and YOLOv8, delivering real-time and accurate road object detection even on hardware-constrained devices.

Keywords

• Road object detection
• YOLOX
• GhostNet
• MCA attention mechanism
• MPDIoU