Indian Traffic Sign Dataset for Self Driving Cars in YOLO Format

This paper presents a new dataset [1] of Indian traffic signs collected along the Mumbai-Pune Expressway and evaluates the performance of the YOLOv8 model in recognizing these signs. The model exhibited strong recognition capabilities for well-represented signs, such as “Ahead Only” and “Bump,” achieving notable levels of recall and precision. However, the study also revealed considerable challenges in identifying Indian traffic signs. A significant issue was the class imbalance in the dataset, particularly for less-represented signs, such as “No Stopping” and “Speed 100,” which adversely affected the model's performance. Furthermore, the detection tasks were complicated by the presence of small bounding box sizes and the model's struggles to differentiate signs from surrounding environmental features, resulting in a moderate mean Average Precision (mAP) of 0.440 at an IoU threshold of 0.5. These results highlight the urgent need for region-specific datasets and suggest future enhancements, including rectifying dataset imbalances through augmentation or synthetic data generation, as well as utilizing advanced structures, such as Feature Pyramid Networks (FPN), to improve the detection of small objects. In conclusion, this research provides critical insights essential for the safe integration of autonomous vehicles in India.