DSG-YOLO: An Efficient Adaptive Attention Network for Real-Time Vehicle Detection in Complex Traffic Scenarios Extended via Decision-Level Radar Fusion

Reliable perception under adverse conditions remains a fundamental challenge for intelligent transportation systems. To address the limitations of vision-based detection in complex traffic scenarios, this article proposes DSG-YOLO, an efficient traffic-scene adaptive attention network, together with a robust decision-level visual-radar fusion strategy. The design explicitly targets four real-world pain points: crowding-induced localization ambiguity, small and occluded objects, multiscale feature mismatch, and inconsistent cross-modal cues. Specifically, a lightweight grouped spatial-aware attention module (GSAM) is introduced to mitigate localization ambiguity in cluttered environments. A spatial and channel synergistic attention (SCSA) module combined with omnidimensional dynamic convolution (ODConv) enhances detection of small and occluded objects via adaptive multiscale fusion. To achieve consistent multimodal association, a decision-level fusion framework based on intersection-over-union (IoU) and global nearest neighbor (GNN) matching is proposed. Experiments on KITTI, RTTS, and real-vehicle platform demonstrate that DSG-YOLO achieves 90.1% mean average precision (mAP) on KITTI and 55.3% on RTTS. On real-vehicle data, the fusion improves precision by 3.11%, reduces false positives (FPs) by 1.61%, and enhances the robustness of real-time physical measurements, demonstrating its potential for intelligent transportation.