GESEI-DETR: A Lightweight Traffic Light Detection Algorithm for Autonomous Driving

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Traffic light detection remains challenging under long-range imaging, illumination variations, partial occlusion, and interference from visually similar light sources such as vehicle headlights and street lamps, which can lead to missed detections, false positives, and localization errors. To address these issues, we propose GESEI-DETR, an improved Real-Time Detection Transformer (RT-DETR)-based traffic light detection algorithm. To mitigate feature degradation in distant small targets while reducing computational complexity, GhostGELAN replaces the ResNet18 backbone. To suppress background interference caused by complex illumination and visually similar light sources, an ECSA feature-enhancement module strengthens informative target responses along both channel and spatial dimensions. To improve multiscale feature propagation and preserve local details, SimAMRepPConv replaces RepC3 in the CCFM, and an EFPN multiscale feature-enhancement module is designed. To improve bounding-box regression for small traffic lights, an InnerWiseMPDIoU loss is constructed. In addition, a cross-task transfer strategy from single-class pretraining on TLD to 11-class color and state detection on CTLD is introduced to strengthen the relationship between global structural priors and fine-grained category features. Finally, LAMP, recovery fine-tuning, and Channel-Wise Distillation (CWD) are employed to further reduce computational complexity. Experimental results show that GESEI-DETR achieves 94.0% precision, 91.7% recall, 94.0% mAP50, and 72.6% mAP50:0.95 on the CTLD test set, representing improvements of 3.1, 5.3, 5.4, and 5.3 percentage points over the baseline, respectively. Meanwhile, the computational cost decreases from 57.0 to 20.4 GFLOPs. These results demonstrate that the proposed method improves traffic light detection performance while substantially reducing computational complexity.