MULTI-SENSOR CLOUD OF DATA FOR PLANT PHENOTYPING AND MOTION ESTIMATION OF AUTOMATED MACHINERY WITH AGRICULTURAL APPLICATIONS

Abstract

Precision Agriculture (PA) appears as one of the key factors to enhance agricultural production to satisfy the world’s increasing demand by providing detailed knowledge of the field and how to treat it accordingly. To do so, PA needs to satisfy three requirements: (i) identify each field location, (ii) collect, interpret and analyze data, and (iii) adjust the use of variable-rate inputs. The Global Navigation Satellite System (GNSS) has become the primary solution for identifying each field location. However, in ground applications, the dense foliage and high vegetation of many orchards environments could make the GNSS antenna readings unreliable due to satellite occlusions. Meanwhile, data is commonly collected using autonomous or semi-autonomous vehicles equipped with a variety of proprioceptive and exteroceptive sensors. Proprioceptive sensors are commonly used to obtain the internal parameters of the vehicle. Those sensors, however, come at the cost of invasive and hard troubleshoot implementations, which is the case of odometers and torque sensors. Conversely, exteroceptive sensors are often more versatile and are used to perceive the surrounding environment. The huge amount of information produced by exteroceptive sensors can also be used, among other things, to estimate parameters associated to the vehicle, such as the motion and the carried weight. The latter is the main issue faced in this thesis, to analyze the use of 3D point cloud information for the estimation of parameters associated with the machinery that can lead to a more accurate and reliable representation of the environment. Through different applications, it is shown the importance of the localization system and how a degraded signal of GNSS influences site-specific applications, specifically those applications based on the characterization of crops. Three different approaches are presented for improving the localization of autonomous vehicles, all of them based on scan matching techniques. The first approach fused a scan matching approach with the degraded GNSS measurements; meanwhile, the second and third approaches consider the scan matching as a self-positioning system. Such approaches are intended to improve the convergence of scan matching techniques by a prior manipulation of the point cloud data. The above is validated through different experimentation using field data.