A Method of Grasping Detection for Kiwifruit Harvesting Robot Based on Deep Learning

Kiwifruit harvesting with robotics can be troublesome due to the clustering feature. The gripper of end effector will easily cause unstable fruit grasping, or bending and separation action interfere neighboring because an inappropriate grasping angle, which further affect success rate. Therefore, predicting correct angle for each guide safely approach, grasp, bend separate fruit. To improve rate rate, this study proposed a detection method kiwifruit robot based on GG-CNN2. Based vertical downward growth characteristics kiwifruit, configuration manipulator was defined. clustered mainly divided into single fruit, linear cluster, other dataset included depth images, color labels. GG-CNN2 improved focal loss prevent algorithm from generating optimal in background at edge performance test network verification robotic picking were carried out orchards. results showed that number parameters 66.7 k, average image calculation speed 58 ms, accuracy 76.0%, ensures run real time. indicated combined position information provided by target YOLO v4 could achieve 88.7% drop 4.8%; time 6.5 s. Compared only provides information, presented advantages when clusters, especially slightly increased. is suitable near-neighbor multi-kiwifruit picking, it harvesting.