Deep Learning-Based Instance Segmentation for Enhanced Navigation of Agricultural Vehicles

• Renato de Avila Lopes UFABC http://orcid.org/0009-0002-8108-1326
• Marcus Vinicius Leal de Carvalho USP https://orcid.org/0000-0002-1298-8517
• Edson Kitani FATEC
• Francisco de Assis Zampirolli UFABC https://orcid.org/0000-0002-7707-1793
• Leopoldo Yoshioka USP
• Luiz Antonio Celiberto Junior UFABC
• Ugo Ibusuki UFABC

Abstract

This paper presents the development of a computer vision application based on the YOLOv8 network, designed to assist the navigation of autonomous vehicles on rural roads, particularly those found in sugarcane fields. The application employs instance segmentation to differentiate between navigable and non-navigable areas and detect obstacles such as pedestrians, vehicles, and other potential hazards. This information is used to generate an occupancy map that helps the navigation planner identify the safest and most efficient routes. The system was trained on a dataset containing 1.018 images, and the results demonstrate that instance segmentation significantly enhances the precision and safety of autonomous navigation in complex rural environments. The proposed approach is compatible with the ROS2 framework, using its structure for data integration and enabling real-time decision making.

Keywords: Instance Segmentation, Yolov8, Computer Vision, Autonomous Vehicle