Real-time in-field trunk size estimation using low-cost RGBD cameras

Trunk size estimation is important to enable crop load management on modern innovative orchards and vineyards. Manual counting and sizing are labour intensive tasks that cannot be done at a scale. Robotics and autonomous vehicles can monitor whole fields/orchards to adopt precision agriculture techniques at a plant/tree level. In this scenario, ground vehicles or tractors equipped with RGBD cameras can estimate the trunk cross sectional area. Once the trunk is detected by a self-trained object detection neural network, a computer vision algorithm determines the trunk boundaries so that a pixel dimension can be reliably computed. Then, using the distance information of the object, it is possible to measure the trunk diameter. The depth information can also be exploited either to compare with the trunk size measurement or to detect irregular trunk shapes. This study used an Intel RealSense D435 depth camera that features an ideal distance range of 0.5-2.0 m, combining colour and depth information. Experiments were carried out at the experimental fields of the University of Bologna, located in Cadriano, Bologna (Italy). Testing was focused on 2D training systems, such as the Guyot used in vineyards and also in apple orchards. The system was mounted on the unmanned ground vehicle from the university spin-off company ‘Fieldrobotics’, featuring autonomous navigation inside orchards. The system was tested on a single row of 160 grape vines and compared to manually sized and georeferenced data. The localization procedure produced an overall RMSE of 0.30 m and the sizing process generally underestimated the trunk diameters. Future improvements will include multiple detection and trunk tracking algorithms to refine the geo-localization and improved segmentation by synchronizing the colour and depth images for diameter estimation.