Object Detection And Tracking Approach To Control Of a Mobile Agriculture Robot

Abstract

The application of machine learning, IoT, and robotics in agriculture increases rapidly during the last years. We are developing a mobile robot for weed destruction. Machine learning and computer vision are an important part of its control system. The objective of this study is to provide research on image recognition and localization of plants and weeds with methods of machine learning, so the development of mechanical and electrical hardware is omitted. This paper presents research on training of the YOLOv5 algorithm for object detection in order to develop a control system of the robot. First a dataset of a cabbage and weed has been created. The dataset consists of three annotated sets of images: training, test and validation. Second, a model is trained and evaluated. Next, bounding boxes with coordinates of weed and cabbage are obtained. Finally, an object tracker is made, it assigns ID to the target cabbage and weeds. In order to control the robot and its end-effector properly both are required the coordinates and ID of each weed and cabbage.