SC-YOLO: Robust Multi-Scale Small Object Detection for Intelligent Transportation

Existing multi-scale vehicle detection methods often falter in crowded traffic scenarios—particularly when it comes to locating small vehicles, resolving occlusions, and adapting to scale variations—which leads to a marked drop in overall accuracy. To overcome these challenges, we introduce SC-YOLO, a lightweight detection framework built upon YOLOv10n and optimized for greater efficiency. First, we replace the standard downsampling between the backbone and neck with a Space-to-Depth Convolution (SPDConv) module, preserving fine-grained details in the lower levels of the feature pyramid so that cues for small targets remain intact. Next, we propose a Context-Guided Rectangular Feature Pyramid Network (CGRFPN) equipped with a self-calibration mechanism; by enabling cross-scale interactions and adaptive feature-map calibration, it significantly enhances multi-scale fusion. Finally, guided by extensive empirical evaluation, we adopt the Wise-IoUv3 dynamic loss function, whose adaptive gradient allocation refines bounding-box regression. On the Pascal VOC, KITTI, and Cars datasets, SC-YOLO attains mAP@50 scores of 79.0%, 87.3%, and 74.6%, respectively—improving upon the YOLOv10n baseline by 2.5%, 2.1%, and 2.3%. Crucially, it maintains high accuracy under challenging traffic conditions, especially for small-vehicle detection and occlusion resolution, while scaling more efficiently, requiring fewer computations than other models with comparable parameter counts. These combined advantages underscore SC-YOLO's resource-efficient design and its practicality for intelligent transportation and autonomous-driving applications.