D-TLDetector: Advancing Traffic Light Detection With a Lightweight Deep Learning Model

Abstract

Traffic signal light detection poses significant challenges in the intelligent driving sector, with high precision and efficiency being crucial for system safety. Advances in deep learning have led to significant improvements in image object detection. However, existing methods continue to struggle with balancing detection speed and accuracy. We propose a lightweight model for traffic light detection that uses a streamlined backbone network and a Low-GD neck architecture. The model’s backbone employs structured reparameterization and lightweight Vision Transformers, using multi-branch and Feed-Forward Network structures to boost informational richness and positional awareness, respectively. The Neck network utilizes the Low-GD structure to enhance the aggregation and integration of multi-scale features, reducing information loss during cross-layer exchanges. We introduce a data augmentation strategy using Stable Diffusion to expand our traffic light dataset in complex weather conditions like fog, rain, and snow, improving model generalization. Our method excels on the YCTL2024 traffic light dataset, achieving a detection speed of 135 FPS and 98.23% accuracy, with only 1.3M model parameters. Testing on the Bosch Small Traffic Lights Dataset confirms the method’s strong generalization capabilities. This suggests that our proposed method can effectively provide accurate and real-time traffic light detection.