|M.Sc Student||Ben Shaul Ron|
|Subject||Examination and Comparison of Mapping and Navigation|
Algorithms for Autonomous Vehicle in Orchards
|Department||Department of Autonomous Systems and Robotics||Supervisor||Professor Amir Degani|
Agricultural production is a central part of the global economy and accompanies humanity from ancient times. By the year 2050, the world's population is expected to grow to about 9 billion and food production will need to increase respectively. In order to deal with this fact, and in addition to the limited resources of appropriate land for growing, water and workforce, we will need to find ways to do that production work more efficiently, accurately and cheaply.
The purpose of this research was to implement and test algorithms for mapping and navigation of an autonomous vehicle in an orchard. The mapping task is done by manually operating the vehicle along the orchard rows. Based on that map, a navigation task to a specific target is performed. While navigating, planning algorithms are performed for selecting the optimal route on the map and to avoid colliding into dynamic obstacles. These capabilities can serve as an extensive basis for the development of automation tasks in orchards such as mowing, spraying and harvesting and enable data collection at the fruit and tree levels for indirect tasks such as disease prediction, insect monitoring and crop yield estimation. The main goal of these tasks is to increase the orchard yields, to bring streamlining and to save resources.
To fulfill the requirements of the orchard mapping task we used the Simultaneous Localization and Mapping algorithm (SLAM) and examined it in several implementations. For the navigation task, we used a global and local planning algorithm that used Monte-Carlo localization algorithm for finding the vehicle position on the map. The SLAM algorithm uses the encoders and the inertial measurement unit sensors to determine the vehicle's progress and direction. In addition to these sensors, the algorithm receives information from a laser scanner that enables it to identify obstacles and insert them into the map. We tested the effect of the orchard structure and environment on the quality of the mapping, and the autonomous navigation task. The research project implements and examines these algorithms in the environment of a physical simulation for a GRIZZLY vehicle. We used the Robot Operating System (ROS) to implement the algorithms since it includes many basic software modules for robotics and a built-in interface between the simulation world and the real GRIZZLY vehicle.
Our conclusion is that using the SLAM algorithm based only on a fast laser scanner (Hector SLAM), ignoring the skid of the wheels and the inaccuracies of the motion sensors, is the most effective and gives the maximum accuracy. This fact is true if there are contiguous zone that can be identified and tracked. As for the localization of the vehicle in the map during the navigation task, the simulations show that as long as the uncertainty in the position of the vehicle is small and limited to a small place, the algorithm manages to converge and shorten the time for the localization of the vehicle. In cases where an evaluation error was entered, an incorrect convergence was observed for the position revaluation.