Road Segmentation with U-Net Architecture Using Jetson AGX Xavier For Autonomous Vehicle

Abstract

Autonomous Vehicle is a technology that has been often discussed in the last few years in the category of research and industry. This technology is able to sense the surrounding environment and control the vehicle autonomously without any human intervention. In its implementation, this technology requires a lot of information, especially the road track that will be passed. Because of that, the thing that must be considered is to segment the road first. The aim of this research is to develop a method that can segment the roads to produce a model that can recognize the road track as well. This research uses Convolutional Neural Network (CNN) with U-Net architecture. The datasets have a form of car trips video recordings from the dashboard camera, which are then extracted into a frame. After this process, it is annotated or manual segmentation using Supervisely to be used as a reference for training and testing. From the results of the calculation process with the confusion matrix, the accuracy of the U-net architecture gets a value of 95%, precision value is 81%, recall value is 92%, F1-Score value is 86% IOU value is 76%. Followed by testing the model in real-time using Jetson AGX Xavier, this tool is specially designed to develop artificial intelligence with high specifications. The test is carried out with two types of testing. The first test with an RGB background produces an FPS of 0.17, and the second test without an RGB background gets an FPS in the range of 0.55-0.67.