LiDAR-Based Row Detection and Model Optimization Design for UAV in Vineyards

During the last few years, the field of agriculture has seen a revolution carried out by the application of new technologies in the field of machine learning, computer vision, automation and decision-making with the goal of not only improving the efficiency and yield of the crops but also reducing costs and environmental impact. This is achieved by analysing data on the crops and terrain in order to determine the best treatment and by designing systems that optimise the delivery of chemicals. With this vision in mind, the Precision Spraying project, developed by the Department of Agricultural, Forest and Food Sciences at the University of Turin, aims at developing a fully autonomous Unmanned Aerial Vehicle (UAV) system capable of performing precision spraying missions in vineyards. The goal of this thesis is double. The first objective is the design of an algorithm capable of identifying the vine row hence determining the reference input for the controller algorithm that will be studied in future work. The second goal is to develop a framework for the modelling and testing of the whole UAV guidance system in a software-in-the-loop simulation and, with this tool, determine an optimal design for the row detection algorithm and Light Detection and Ranging (LiDAR). In order to efficiently perform precision spraying the UAV position with respect to the crop that needs to receive the treatment is crucial for maximising the amount of plant reached by the chemical while minimising the amount of product wasted on the ground, how well the drone can maintain this optimal position is greatly affected by the performances of the row-detection algorithm and for this reason, in this thesis, three different methods are analysed: RANSAC, Hough Transform and Iterative Skewing. Due to the costs derived by the production of a functioning prototype in this thesis, a model-based design approach is investigated in order to determine how well commercial LiDAR sensors perform with the different algorithms and in different conditions. For this purpose, a performance evaluation tool is proposed. This evaluation tool analyses the performances of the row-detection algorithm and of the LiDAR in many different conditions thanks to a simulator that, given the current state of the drone, generates the LiDAR sensor's corresponding output as a point cloud.