Improving Traffic Sign Detection Under Variable Lighting Conditions Using Deep Learning Algorithms
P. Nithya, V. Sowmitha, Malatthi Sivasundaram, V. Lokeshwaran, V. Gopinath, R. Latha · 2025
Aim: In this research, the accuracy of traffic sign recognition in varying lighting conditions is enhanced by utilizing deep learning methods such as ResNet50, Deep Convolutional Neural Networks (DCNN), and YOLO v8. The study targets the detection of both symbol-based and text-based traffic signs with strong recognition under various lighting conditions. Materials and Methods: The dataset for this study includes traffic sign images from the European Traffic Sign Dataset (ETSD) and an extended German Traffic Sign Detection Benchmark (GTSDB). A Kaggle.com dataset was utilized, and feature extraction methods such as Local Binary Pattern (LBP) and Mean-Median Variance. Two groups were compared: Group 1 (Existing Methods - DCNN & ResNet-50) having 5000 images in its sample size and parameters such as accuracy, false positive rate, and recall; and Group 2 (Proposed Method - YOLO v8) also having 5000 images, assessed on detection time and accuracy. Results: It indicates that YOLO v8 performed at 95% accuracy, which is better compared to DCNN (85%) and ResNet-50 (88%), with a reduced error rate of 4.2%. Conclusion: The research verifies the better performance of YOLO v8 and makes it a potential real-time traffic sign detection solution. Such results indicate the potential of deep learning to advance road safety and navigation management.