Content area
Full text
Abstract. The paper presents a method of terrain classification and path planning for unmanned ground vehicles. The terrain classification is done on imagery that is acquired from UAV (unmanned aerial vehicle) or satellite and is used for UGV (unmanned ground vehicle) path planning thus introducing collaboration capabilities to the system of two. The system complements the UGV on-board navigation system by increasing its perception distance and providing longrange path planning capability.
Key words: optical terrain classification, UAV, UGV, path planning, convolutional neural networks.
(ProQuest: ... denotes formula omitted.)
1. INTRODUCTION
Our ultimate target is to build an UGV that is capable of driving independently to given GPS coordinates. For off-road navigation it is important to know what is behind a bush or a house ahead for cul-de-sac or rough terrain avoidance. Having fresh data about terrain behind horizon helps to reduce time and energy spent on wandering and to avoid potentially dangerous terrain that is hard to detect on time with on-board sensors (ditches and cliffs). The perception distance of an UGV is usually limited to visibility range of its on-board sensors; it is rarely over 100 m [1], which is sufficient for obstacle avoidance but is insufficient for long-term path planning.
To increase UGV perception distance and overall performance we use aerial imagery provided by UAV for analysing terrain ahead of the UGV and for generating path to the target position. We detect a set of features on aerial imagery that should be preferred (such as roads, grass) or avoided (buildings, water) during navigation and feed the acquired information into the path planner.
Effectiveness of fusing UGV on-board sensor data with aerial imagery has been previously demonstrated in [2]; in these experiments with aerial data significant increase in UGV average speed and decrease in required human interventions has been measured. These experiments, however, relied heavily on 3D point cloud acquired by LiDAR mounted to UAV but our system must be limited to passive sensors (cameras, gyro, GPS).
Suitability of convolutional neural networks for terrain classification was demonstrated by Sermanet et al. [3,4]; they built a robust UGV navigation system that relied solely on visual data. Their system uses two-level architecture: fast obstacle detection module with perception range of around 5 m and...