Harvest Crate Detection for Grapes Harvesting Robot Based on YOLOv3 Model

摘要

Fast and accurate algorithms for object detection allow robotic systems to real-time process scene information identically to the human visual system. The robotic system of our interest is a grape harvesting robot. The robot will cut the grape bunches suitable for harvesting and collect them in harvest crates located in the vineyard corridors along its path. In this work, a state-of-the-art algorithm, the You-Only-LookOnce version 3 (YOLOv3), is employed to solve the in-field demanding object detection task, that of harvest crates. Complicated environment in-field conditions, such as illumination variations, branches and leaves occlusion, weed and object overlapping, make harvest crates detection challenging. The proposed algorithm is suitable to deal with these problems. Experimental results verify the effectiveness of the model under varying conditions, providing high recognition accuracy up to 99.74 % (mAP). Thus, the model is able to provide reliable information to the harvesting robot, which is the required preliminary step so that the robot could navigate to the harvest crate and place the grape bunches.