Multi-View Crop Detection Using an Occupancy Grid for Agricultural Applications

To progress towards sustainable agriculture, crop production processes have to be optimized. Avoiding the use of herbicides is a major issue and a way to go on the agro-ecological transition; hence the interest in robustly differentiating crops from weeds and suppressing the latter. We present in this paper a multi-view crop detection methodology to locate and track crops; allowing a mobile platform to remove weeds with high precision. The mobile platform is endowed with two perception systems, fixed at two different positions and orientations with respect to its center of mass, providing two different points of view of the crops. The proposed methodology merges both fields of view in a single occupancy grid to precisely locate crops in front and below the mobile platform. Each perception system is an edge-computing device capable of preprocessing concurrently raw imagery, facilitating the task of merging both points of view. To validate the proposed approach, multiple simulations were carried out using the Gazebo simulator. In the same manner, we also tested this methodology in a real agricultural field during the evaluation campaigns of the ROSE challenge.