Precision Evaluation of GPS based Autonomous Agricultural Vehicles using Computer Vision

Technological advances have been successfully achieved in precision agriculture using autonomous agricultural vehicles. Among these advances, the increase of efficiency and productivity in field operations can be highlighted. Several autonomous driving systems are implemented using the GPS RTK system, which allows operations to centrimetric accuracy. However, irregularities in ground conditions, tractor traction, wheel slip and operating speed may influence the performance of GPS based autonomous agricultural vehicles. In this way, the evaluation of the autonomous driving systems becomes essential to the achievement of high precision levels in field operations. This evaluation can be performed by measuring the displacements using locally installed sensors in the vehicle, such as: cameras, lasers, odometer, ultrasonic sensors, among others. Among the local sensing options, it is well-know that computer vision methods allow the location of any system in the space. Nevertheless, these methods demand the adjustments of their parameters to ensure high accuracy. In this way, the objective of this work is to evaluate the precision of an agricultural vehicle in an autonomous condition using computational vision methods and image processing techniques. Tracking localization by matching key points in digital images can be exploited in order to assess the location of the vehicle during its work in the field. The outcome of this proposal can be evaluated to infer conclusions about the accuracy of the autopilot system. The vehicle under study is a Massey Ferguson 7350 with the Auto-Guide 3000 autopilot system with GPS RTK correction signal. The computer vision system consists of two Canon Rebel T5 cameras with focal lens of 50 millimeters. The image processing was performed using a corners’ detector technique developed in a grid image in the field. The manuscript details the camera’s calibration and the vehicle’s localization procedures. The main conclusion of this work is that computer vision can be successfully exploited for aiding the autonomous driving of agricultural vehicles if devices, procedures and parameters are well selected.