A real-time and lightweight vehicle detection algorithm with multi-scale feature fusion based on you only look once

Accurate and efficient vehicle detection is a cornerstone of intelligent transportation systems and autonomous driving. Nevertheless, challenges such as small-scale targets, occlusions, and complex traffic scenes often compromise detection performance. To address these issues, this paper introduces You Only Look Once Vehicle Detection (YOLO-VD), a lightweight real-time framework with enhanced multi-scale feature fusion. Specifically, a Cross Stage Partial Efficient Attention Network (CSP-EAN) incorporating the proposed Attention-Driven Convolution (ADConv) is designed to strengthen feature representation and inter-channel interactions. A Grouped Spatial Pyramid Pooling-Fast (GSPPF) module, equipped with Grouped Lightweight Spatial Kernel Attention (GLSKA), is further employed to efficiently capture multi-scale contextual dependencies. Moreover, an additional P2-level detection head is incorporated to enhance small-object recognition, and an EffilightHead is developed to refine channel interactions via adaptive group convolution and feature recalibration. Extensive experiments on BDD100K and SODA10M demonstrate that YOLO-VD achieves precision, recall, mean average precision (mAP), and frames per second (FPS) of 64.3%, 50.7%, 55.2%, and 51.1 on BDD100K, and 66.1%, 43.5%, 49.0%, and 47.3 on SODA10M, respectively. Compared with the YOLOv11n baseline, YOLO-VD improves mAP by 8.0% and 6.7%, while maintaining high inference efficiency (78.9 FPS and 68.4 FPS with individual modules). These results highlight YOLO-VD's capability to balance accuracy, efficiency, and scalability, underscoring its potential for real-world traffic perception tasks.