ICWT 2026: Improving YOLOv12 with Selective Kernel Networks for Autonomous Driving Object Detection

The 12th International Conference on Wireless and Telematics (ICWT) 2026 "Improving YOLOv12 with Selective Kernel Networks for Autonomous Driving Object Detection" Pandu Pangestu, et al. Abstract: Object detection in autonomous vehicles requires a model that can preserve accuracy under object-size variation caused by camera distance and dynamic road environments. YOLOv12 provides a real-time architecture strengthened by attention mechanisms, but its multiscale feature representation can still be improved through adaptive receptive-field selection. This paper proposes the integration of Selective Kernel Networks into YOLOv12n to improve feature representation for autonomous driving object detection. The SKNet module is inserted into the feature aggregation path to dynamically select kernel contributions through split, fuse, and select operations. The evaluation is conducted on the KITTI dataset with eight object classes using precision, recall, mAP, parameter count, GFLOPs, and inference time. Experimental results show that YOLOv12n with SKNet improves precision from 90.1% to 92.8%, recall from 76.0% to 83.0%, mAP50 from 87.0% to 91.3%, and mAP50-95 from 64.3% to 70.0%. This improvement is accompanied by an increase in parameters from 2.51M to 2.88M, GFLOPs from 5.8 to 7.7, while the inference time remains 1.6 ms per image. These results indicate that SKNet improves YOLOv12n detection performance with measurable computational overhead. Keywords: Autonomous Vehicles, Deep Learning, KITTI dataset, Object Detection, Selective Kernel Network, YOLOv12