An Omni-Dimensional Dynamic Convolution-Based Object Detection Algorithm for Autonomous Driving

Abstract In this study, an autonomous driving object detection algorithm named ODC-YOLO (Omni-Dimensional Dynamic Convolution–You Only Look Once) is proposed based on Omni-Dimensional Dynamic Convolution (ODConv). The KITTI 2D object detection dataset is used to train and evaluate the ODC-YOLO model. Compared with traditional autonomous driving object detection methods, the main novelty of ODC-YOLO is its use of Omni-Dimensional Dynamic Convolution and an improved decoupled detection head to enhance feature extraction and facilitate model convergence, enabling more accurate detection of various targets in autonomous driving scenarios. Furthermore, the original CIoU loss was replaced with EIoU, and a Focal weighting mechanism was incorporated to adaptively reweight bounding-box regression samples according to their localization quality. Experimental results show that ODC-YOLO achieves an [email protected] of 87.1%, representing absolute increases of 6.7% and 1.2% over the [email protected] values of YOLOv5 and YOLOv8, respectively. These results demonstrate the effectiveness of the proposed method for object detection in the road scenes represented by the KITTI dataset.