Automatic Detection and Categorization of Road Traffic Signs using a Knowledge-Assisted Method
Vivek Srivastava, Sumita Mishra, Nishu Gupta
Abstract
Deep Learning is proving to be a boon for the latest smart city applications. For efficient mobility-based applications, Artificial Intelligence (AI)-based Intelligent Transportation Systems (ITS) are indispensable. In an autonomous driving scenario, a Traffic Sign Detection and Recognition system (TSDR) helps the drivers. Eventually, with its proper use, the number of accidents on the road can also be reduced. TSDR is an important technique in ITS but it can be hard to implement because of the mess, different levels of light, different sign sizes, and changing weather. In the past, different object detection models such as Fast regions with convolutional neural networks (RCNN) and Faster R-CNN were utilised for this purpose but due to the sluggish pace of detection, these models are inefficient. In this article, a state-of-the-art approach using ”You Only Look Once” version 4 (YOLOv4) is being used for the correct detection and recognition of traffic signs. YOLOv4 utilizes convolutional neural networks (CNN) for the instant detection of objects. YOLO framework processes input image in a single instance and it does not require a complex detection pipeline, making it extremely fast. Therefore, the proposed approach is feasible for real-time traffic sign detection and recognition. The detection accuracy of the proposed YOLOv4 model is improved by adjusting the relevant parameters and retraining the model on manually annotated data set. It has been demonstrated in results that the performance of proposed model is adequate.