TS-YOLO: a lightweight real-time traffic sign detector in resource-constrained environments
Hongxia Yu, Wenhui Liu, Tianhua Xu · PeerJ Computer Science · 2025
Traffic sign detection is essential for ensuring driving safety in autonomous vehicles and advanced driver-assistance systems, particularly under limited computational resources and challenging real-world conditions such as extreme weather, occlusion, and diverse viewing angles. To address the trade-off between detection accuracy and computational efficiency, we propose TS-YOLO, a lightweight, real-time detector built on You Only Look Once version 8 (YOLOv8). TS-YOLO integrates a Repetition Cross Stage Partial (RepCSP) feature-enhancement module and a set of targeted training strategies designed to improve small-object recognition and to handle various challenging scenarios. Experimental results on the CCTSDB2021 dataset show that TS-YOLO achieves 98.4% mean Average Precision (mAP)@0.5 and 79.1% [email protected], with an inference time of only 0.6 ms per frame on embedded hardware. Compared to state-of-the-art methods—including Faster Region-based Convolutional Neural Network (Faster R-CNN), Efficient Object Detection (EfficientDet), and the original YOLOv8—TS-YOLO demonstrates superior accuracy and faster inference. Further robustness evaluations confirm consistent performance across adverse weather (fog, rain, low light), various road types (urban, rural, unstructured), and challenging object characteristics (scale variation, occlusion, irregular shapes). These findings suggest that TS-YOLO is a practical, high-performance solution for traffic sign detection in resource-constrained, complex driving environments.