基于 3D 视觉的林果采摘最佳观测点预测方法研究

Aiming at the problem of poor target localization accuracy caused by severe occlusion of forest fruits in unstructured agricultural scenarios, a 3D active vision-based optimal observation position prediction method for forest fruit picking was proposed, with a six degrees of freedom (6-DOF) robotic arm as execution body. The method mainly consisted of an active vision perception module, and the active vision perception module first accurately identified forest fruits through image processing methods and obtained occlusion status of the fruits. Then, taking the current observation position as starting point, the robotic arm was moved within a certain range to generate a confidence and occlusion rate distribution map. After a series of occlusion rate and confidence datasets were established, a deep learning network was designed to predict the pose with the minimum occlusion rate under current occlusion condition, i.e., the optimal observation position. To verify the effectiveness of this method in improving the accuracy of target localization, field fruit picking experiments were conducted. After 100 trials, results show that the picking success rate of this method is 65%, which is 32% and 24% higher than the binocular camera method and the simple teaching method, respectively.