S 2 -DETR: Hierarchical Sparse-to-Spatial Attention Enhanced DETR for Traffic Participant Detection in Sparse Autonomous Driving Scene

Multi-scale detection is vital for autonomous driving. In sparse scenarios such as highways, where targets often appear small, distant, and susceptible to substantial background interference, traditional models suffer from feature distortion, leading to missed detections and compromised safety. This study proposes S2-DETR to address these challenges, featuring an innovative hierarchical Sparse-to-Spatial Attention Mechanism (S2AM), which synergistically integrates a Dynamic Sparse Attention Module (DSAM) for coarse-grained sparse feature enhancement with a Spatial Attention Module (SAM) for fine-grained refinement. This design is particularly effective for enhancing the representation of small and hard-to-detect targets in complex visual environments. We further designed a Cross-scale Attention Pyramid Module (CAPM) that embeds S2AM within a dual-path architecture inspired by Feature Pyramid Networks and Path Aggregation Networks, replacing RT-DETR’s original fusion module to optimize multi-scale feature representation. Extensive ablation studies validated our S2AM and CAPM designs. Comparative experiments confirmed S2-DETR’s superiority: on public and self-built sparse datasets, it achieved accuracy improvements of 8.7% and 14.1%, respectively, over its RT-DETR baseline, with only a 7.5% speed trade-off. These results establish an improved accuracy-speed balance, notably for challenging small and multi-scale targets. The source code will be released on GitHub to foster further research in traffic participant detection for autonomous driving.