Autonomous Navigation Route Extraction in Dwarf Close-Planted Orchards Based on Improved YOLOv8

This paper presents a method for tree trunk recognition and navigation track line extraction of orchard mobile robot based on improved YOLOv8 model. Firstly, the C2f module of the original YOLOv8 model is replaced with CReToNeXt structure to improve feature sensitivity and reduce computational complexity. Meanwhile, Shuffled-attention (SA) module is introduced to solve background interference and trunk morphological complexity problems. The improved YOLOv8 algorithm was used to identify the tree trunk feature points of the apple tree rows on both sides, and then linearly fit the feature points on both sides through the least square method to obtain the track lines of the tree rows on both sides. Finally, the navigation path of the mobile robot was determined by the reference track lines of the tree rows on both sides. Compared to previous versions, the improved YOLOv8 model shows a marked increase in accuracy in tree trunk detection reaches 90.8% and the accuracy of navigation line extraction reaches 90%, which can effectively extract the navigation path of the orchard mobile robot.