A REAL-TIME TRAFFIC SIGN DETECTION SYSTEM USING A TRANSFER LEARNING APPROACH

This study presents the design and implementation of a real-time traffic sign detection system using a transfer learning approach based on the YOLOv8s model. The system was constructed in a manner that it identified two important road signs which included checkpoint ahead and cattle crossing in diverse environmental conditions including daytime, night and weather conditions. The data set of 640 traffic sign images labelled by the Kaggle repository had been pre-processed with resizing, annotation, normalising, and converting to YOLO format before a 70-20-10 train-test-validation split. YOLOv8s is its architecture based on a CSPDarkNet53 backbone, Efficient Layer Aggregation Network (ELAN), Feature Pyramid Network (FPN), Path Aggregation Network (PAN), and decoupled detection head that was additionally trained using Bayesian hyperparameter optimization to improve accuracy and decrease overfitting. In the experimental results, the system realised incremental performance gains over the training epochs, with the steepest performance improvements realised between epochs 10 and 30. A model achieved a precision of 0.95, a recall of 0.92 and a mean Average Precision(mAP5095) of 0.83 at the final epoch (epoch 50). Performance analysis in terms of classes showed that both category of traffic signs were identified with a high precision with the checkpoint ahead sign slightly doing better than the cattle crossing sign. These findings represent the efficiency, precision, and the computing power of the YOLOv8s model, which proves that this model is highly applicable in the framework of real-time driver assistance and smart traffic surveillance applications. This paper concludes that the presented system is a credible solution that can be used to detect traffic signs and implement it in a real-life road situation.