Autonomous navigation method of jujube catch-and-shake harvesting robot based on convolutional neural networks

Abstract

Autonomous navigation of agricultural devices plays an indispensable step for the realization of various field intelligent operation tasks. However, the accurate extraction of navigation path is still a challenging task due to the complex orchard environment with many interference factors. To cope with the challenge, a low-cost vision system based on convolutional neural network and machine learning algorithm is proposed to obtain accurate navigation path for autonomous path detection of catch-and-shake harvesting robot under winter jujube orchard. In the vision system, firstly, an improved light-weight YOLOX-Nano architecture with some tricks including CSP attention block, SPPCSPC-F and ASFF is proposed to detect root points of jujube tree. Results show that the mAP of our model is 84.08 % and model size is only 12.30 MB, which can meet the requirements of embedded deployment. Secondly, a K-means clustering algorithm is applied to divide root points into two classes which belong to left and right tree row. Finally, the navigation line can be determined by geometric relations and least squares. The experimental results showed that the navigation centerline mean pixel errors and average heading angle deviation are 4.49 pixels and 2.55°, respectively. Therefore, this method can provide a new theoretical basis and technical reference for autonomous navigation of winter jujube catch-and-shake harvesting robots.