Deep Learning Approach for Traffic Sign Recognition on Embedded Systems

Abstract

This research focuses on creating an affordable and effective warning system for drivers that is able to detect the warning signboards and speed limits in front of the moving vehicle and prompt the driver to lower to safer speeds if required. The software internally works on a deep learning-based modern neural network YOLO (You Only Look Once) with certain modifications that allow it to detect the road signs really quickly and accurately on low-powered Advanced RISC Machines Central Processing Units (ARM CPUs). Along with YOLO, the system uses a darknet classifier for second-stage detection. The system is also capable of noting the traffic violations made by the driver like driving over the speed limits prescribed by the Indian Roads Congress (IRC). The system works on the Raspberry Pi board and the Raspberry Pi camera that allows the system to be budget-friendly yet very effective. The dataset for Indian roads is not available in the public domain; hence, majority of the dataset has been created by us, while a small part of the dataset has been salvaged from the Belgium and German traffic sign datasets. Creating the dataset on our own means that images need to be taken of multiple signboards from multiple angles and varying distances. This is a very tedious and long task that is very time-consuming. The first stage involves using a tiny version of YOLO to detect the signboard in a natural scene, in case a speed limit sign is detected by the first-stage detector, the extracted region of interest (ROI) is passed on to the darknet classifier which then accurately classifies the speed limit value on that board.