Trajectory detection in furrows to automate the navigation of Agrobot with AI

Robotics is a technology that is used in various fields; one of these in agriculture, where the robots are known as agrobots. Bearing in mind that in countries like Colombia, the cultivated lands are irregular, most of the time in mountains, and the furrows (which follow paths with very irregular shapes), it is very difficult to use automated machinery to support the cultivation work. Therefore, it is very difficult to use ground or aerial robots to support tasks such as planting, eliminating weeds, detecting diseases , the lack of fertilizers or the presence of water stress. In developed countries, where crops are grown using straight furrows and flat terrain, it is easy to have assistant robots for agricultural work, but in Colombia, where geography imposes irregular terrain, it is not easy and therefore growers cannot have technologies that support their daily work. In order to address this problem, and facilitate the use of agrobots in different crops, mainly potatoes and vegetables, research has been carried out at the Colombian School of Engineering Julio Garavito to propose solutions that give robots autonomy on the ground and air for exploration of crops. One of these works is related to the real-time processing of images, captured by an RGB camera placed on the front of a rover, and processed with deep learning techniques, to establish the correctness of the route that a robot must follow when running furrows in potato crops. Initially, a previous image processing is done to deliver a binary image with the shape of the furrows to a neural network. Said network analyzes the image and determines the correction angle that the robot’s direction must take. For this training, a dataset of 1,000 images per class was created, taken from simulated furrows with artificial plants. The classes used correspond to a discrete scale at intervals of 10 degrees. The results show a degree of accuracy greater than 90% in the automatic detection of the angle that the robot must correct to continue moving over the furrows. In this project, different deep learning and image processing methods were analyzed and tested to give the robot autonomous movement over the furrows.