Indian Traffic Sign Recognition as a Two Stag Approach using YOLOv4

In real-world applications such as autonomous driving and driver safety, Traffic Sign Recognition (TSR) systems play a critical role. In a country like India, where new technology is quickly adopted, automatic vehicles will adapt quickly. Even significant corporate in India Tata Motors is working on bringing self-driving cars to India. Tesla, the biggest developers of selfdriven cars is granted the right to sell cars in India. According to reports, Tesla will be selling cars in India by the end of 2022. The state-of-the-art of object detection models such as YOLOv3, Tiny-YOLOv4 and YOLOv4 is examined in this work. Transfer learning is used to tailor these models to the domain of traffic sign detection. As we are focusing on India, we finetune these state-of-art object detection model trained on COCO dataset using Indian Traffic Sign Detection (ITSD) Dataset. There is no such database, hence this work contribute a new Indian Traffic Sign Detection dataset obtained from random Indian road scenes. The task of recognition is made up of two parts. Detection and classification are two steps in the process. In this work, detection module is implemented using three different YOLO (You Look Only Once) versions and classification using custom Convolutional Neural Network (CNN) model. The German Traffic Sign Detection Benchmark (GTSDB) dataset is also utilized to evaluate the system because it is the most widely used dataset. With a threshold of 0.5, the detectors made with YOLOv4, YOLOv3, and Tiny-YOLOv4 were tested on the ITSD dataset and achieved mean average precision (mAP) of 99.35%, 97.20%, and 80.89%, respectively.