A Lightweight Traffic Light Detection and Recognition Model Based on YOLOv5

With the rapid development of intelligent mobile platforms, real-time and accurate Traffic Light Detection and Recognition (TLDR) is critical for ensuring traffic safety. However, existing TLDR methods fail to balance detection accuracy, speed, and deployment feasibility. To address this, we propose a high-precision lightweight TLDR model. First, an enhanced backbone network based on the EfficientViT architecture is proposed, incorporating an efficient multi-scale attention module to extract rich semantic features while reducing computational load. Second, a lightweight cross-scale feature fusion network is constructed to enhance information interaction between deep and shallow features. Finally, the model is trained and validated on the open-source S2TLD dataset. Experimental results demonstrate that the proposed model achieves a detection speed of 204 FPS, precision of 95.0%, and 2.1 GFLOPs. Compared to YOLOv5n, the proposed model improves detection speed by 35.1%, precision by 4.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, and GFLOPs by 48.8%, respectively. This work provides technical support for constructing lightweight TLDR models and offers theoretical reference for deploying TLDR models on intelligent mobile platforms.