Research On Navigation System For Orchard Crop Protection Machinery Based On Multi-Sensor Fusion

Automated driving of orchard crop protection machinery is the core technology for orchard machinery intelligence.Under the limited computational capabilities of embedded systems used in orchard machinery,improving environmental perception and navigation accuracy is a prerequisite for achieving autonomous operations.Due to the potential signal loss of GNSS sensors in the presence of tree obstruction,the influence of soil and weather on camera image data quality,and the high data volume and cost of high-beam Li DAR,a single sensor alone cannot achieve precise environmental perception in orchards.To address these issues,this thesis conducts research on environmental perception and autonomous walking methods for orchard crop protection machinery based on multi-sensor fusion methods.Multiple fruit tree varieties are used as experimental objects,and the effectiveness of the proposed methods is validated through the establishment of datasets,algorithm design,comparison of different algorithm performance,and field experiments.The main content is as follows:(1)Considering the characteristics of large undulations in the orchard terrain,numerous obstacles and obstructions,frequent changes in lighting conditions,sensor occlusion due to soil,and low computational power of orchard crop protection machinery,cameras,Li DAR,and IMU sensors are selected.Zhang Zhengyou’s calibration method is used to calibrate and jointly calibrate the sensors based on coordinate system transformations.Decision-level fusion algorithms are employed to enhance target detection performance and system robustness,thus improving operational accuracy.(2)The YOLOv7 algorithm was improved by introducing attention mechanism and improving bounding box loss function.The accuracy rate is improved by 1.67%,recall rate by 3%,and m AP by 2.97% at a detection speed of more than 50 FPS.Using density-based lidar point cloud clustering target detection algorithm,the target detection effect can be improved in this experimental environment,and the limitation of single sensor in orchard environment is verified;(3)Local path planning is carried out based on the laser radar coordinate system,and the navigation points are extracted by the target point coordinates,so that the problem of difficult extraction of the visual algorithm is solved while reducing the calculation amount.The object and obstacle are distinguished by clustering point cloud volume features,the obstacle avoidance path is planned by Bezier curve,and the inflection point path is smoothed.Different obstacles detection and obstacle avoidance path planning are realized,and the fast path tracking control is realized by using the sliding mode control algorithm.(4)A tracked self-propelled thermal fogger was used as an experimental platform,and a software control framework was developed based on the Robot Operating System.The experiment was conducted in a tree plantation with a row spacing of 3.2 meters,at a speed of 0.6 meters per second.The maximum deviation of the operational path was 17.58 cm,and the average deviation was 8.575 cm,with a variance of 15.39 cm2.These results show that the precision requirements for automated driving of agricultural machinery in conventional orchard environments are satisfied.