YOLOv11-ASD: Improved YOLOv11 for Multi-Scale and Small Object Detection

Object detection plays a vital role in autonomous driving, where achieving both high precision and computational efficiency is a major challenge. To address the limitations of YOLOv11 in feature representation and computational cost, we propose YOLOv11-ASD, an enhanced architecture integrating three lightweight modules: ADown, SPPF-SENetV2, and DySample. ADown optimizes downsampling by reducing redundant convolutional operations; SPPF-SENetV2 enhances global receptive fields and channel-wise attention; DySample introduces a dynamic point-based upsampling mechanism at lower computational cost. Experiments on the KITTI dataset show that, compared to the original YOLOv11, YOLOv11-ASD improves [email protected] by 2.4% (reaching 93.5%), reduces parameter size 19%, and lowers GFLOPs by 19.2%. The results demonstrate that the proposed model achieves higher detection accuracy and faster inference, making it highly effective for realtime object detection applications.