IMPLEMENTATION OF TRAFFIC SIGNS DETECTION USING YOLOv4 ON JETSON TX2

Abstract

Traffic signs and traffic lights detection is an important aspects of autonomous vehicles to prevent and reduce accidents. In this paper, the YOLOv4 model, one of the popular architectures of deep learning used to detect and recognize traffic signs and traffic lights has been implemented on NVIDIA Jetson TX2 hardware. In general, the YOLOv4 model has many variants with different structures and parameters, the author focuses on comparison and evaluation to select the right variant for the data set. Specifically, the author chooses NVIDIA's Jetson TX2 as a hardware platform to take advantage of the GPU's computing power to optimize data training time. In particular, the usage data is generated independently including 32 classes for training and testing. The method used achieved 91% mAP and 31.2 FPS on the test dataset.