<p>Object detection technology has become an extremely important part of autonomous driving. This presents a significant challenge for algorithms to accurately and quickly identify traffic objects. Therefore, the balance between the efficiency and effectiveness of the object detector is crucial for autonomous driving. To address this issue, we propose YOLO-OAD, an optimized lightweight object detector tailored for autonomous driving. YOLO-OAD introduces: (1) a Deformable Lightweight Asymmetric Decoupled Head (DLADH) to dynamically sample features and improve the efficiency of the detection head, (2) a C3 Large Kernel StarNet Context Anchor Attention (C3LKSCAA) module to obtain non-linear high-dimensional mapping, long-range context dependency, and larger receptive field, (3) a C3 Visual Geometry Group with the simple and Parameter-Free Attention Module (C3VGGAM) to retain more salient semantic information, and (4) a Focal-EIoU loss function to enhance bounding box regression. Experiments show that the YOLO-OAD outperforms the YOLOv5n, improving the mAP@0.5 by 5.4% and the mAP@0.5-0.95 by 8.8% on PASCAL VOC dataset. Further validation on the BDD dataset shows that the YOLO-OAD improves the mAP@0.5 by 5.3% and the mAP@0.5-0.95 by 3.8% compared to the YOLOv5n, with notable gains in detecting motor (+9.4%), person (+6.5%), and rider (+5.3%). In addition, our method runs at 126 FPS, offering a better accuracy-speed trade-off for autonomous driving. The code is available at <a href="https://github.com/clouduy/YOLO-OAD/tree/master">https://github.com/clouduy/YOLO-OAD/tree/master</a>.</p>