Research On Orchard Obstacle Detection Based On YOLO Algorithm

China is a major fruit-producing country.To alleviate the labor intensity and improve efficiency,the automation and intelligence of agricultural machinery equipment in orchards have become a top priority for high-quality development.To ensure safe operations,orchard agricultural machinery equipment must have real-time obstacle detection capabilities.This study employs an improved deep learning object detection method combined with binocular vision to enhance the accuracy and real-time performance of obstacle recognition and positioning in orchard agricultural machinery equipment,addressing the issues of insufficient computational power in embedded computers.The main research content and innovations of this paper are as follows:(1)Constructing a visual perception system on the orchard spraying robot,the binocular cameras are used to capture RGB image and depth information.These two types of information are transmitted to the visual perception system,where the embedded computer performs object detection and ranging to accurately locate obstacles in the orchard.(2)To enable the target detection model on an embedded computer platform while meeting the requirements of real-time detection and accuracy,the Ghost-YOLOv4 detection algorithm is adopted.It addresses the obstacle positioning issue by combining it with the binocular vision camera.The Ghost-YOLOv4 detection algorithm replaces the original YOLOv4 backbone network with Ghostnet and uses the Ghost Module to replace the regular convolutions in other parts,significantly reducing the computational load while maintaining detection accuracy.The EIo U improvement is applied to calculate the interaction degree between the ground truth boxes and predicted boxes in the loss part.Experimental results show that the Ghost-YOLOv4 detection algorithm improves the detection speed by 59.3% compared to the original YOLOv4 detection algorithm,reduces the model size by 89.4%,and maintains almost the same detection accuracy.By estimating the depth information of obstacles using a depth camera,the ranging errors for fruit trees,pedestrians,and pole-like objects within a 5m range during routine operations of the orchard robot are 9.42%,2.10%,and 1.06%,respectively.(3)Addressing the issues of low detection accuracy and the occurrence of false negatives and false positives caused by target occlusion in the YOLOv5 detection algorithm,improvements are made as follows: the CA attention mechanism is introduced at the connection between the backbone network and other parts to incorporate spatial positional information into the channel dimension,enhancing the focus on detection targets.The DIo U NMS algorithm is modified to the Soft-EIo U-NMS algorithm,which uses a Gaussian-weighted function to adjust the object label confidence of the remaining detection boxes instead of direct suppression,reducing false negatives and false positives for overlapping objects.Experimental results demonstrate that the improved YOLOv5 detection algorithm achieves higher accuracy,recall rate,and average precision for all categories by 2.08,1.56,and 2.31 percentage points,respectively.