Off-Road Terrain Identification And Analysis

Abstract

Background: The role of the Terrain is paramount for any autonomous vehicle to drive safely on any type of surface. The Autonomous vehicles should have the capability of identifying the terrain and should adapt to the environment. With the evolution of robotics and Artificial Intelligence, and understanding diverse terrains, the techniques for terrain identification are also advancing with a major focus on safety.
 Methodology: To make Terrain Detection and Identification more reliable we used instance segmentation which is a more sophisticated type of segmentation that goes a step ahead of semantic segmentation by performing both object detection and segmentation at the same time. In order to perform Instance segmentation, we used the YOLOv8 architecture which is considered to be the state-of-the-art CNN (Convolutional Neural Network) architecture. The YOLOv8 model was trained on an Off-road Terrain Dataset.
 Results: Our findings indicate that the state-of-the-art YOLOv8 instance segmentation model provided the best results for terrain detection and segmentation with a threshold confidence of 0.60, and the results provide a maximum confidence of 0.92 which indicates an accurate segmentation model for the given terrain detection problem.
 Conclusion: The present work motivates for a more viable hardware model that makes use of trained computer vision models and cutting-edge sensors that can be tested on different soils and terrain. The results obtained can be used to study about the different Terrains and select the most suitable model, this in turn drives for further research in the subject of Terrain Identification and Detection.