| Given low accuracy,slow recognition speed,and low recognition efficiency of the edge device deployment of recognition model in the existing fruit recognition method in the orchard environment.Relying on the key research and development plan(modern agriculture)project of Jiangsu Province,this paper takes citrus fruits as the research object,and studies the detection and localization of citrus fruits in orchard environment,movement trajectory planning of picking robotic arm,to improve the level of agricultural automation and informatization.The main research works of this paper were as follows:(1)A detection model of citrus fruits was constructed.The one-stage detection YOLOv4 model was used as the base model for citrus fruit detection,and improved the backbone network based on SENet attention module to improve the sensitivity of the model to useful features;CSPConv was designed to improve the neck network structure to reduce the computation while enhancing the reuse of the underlying features;the confidence loss function based on Focal loss was reconstructed to improve the detection ability of the model for indistinguishable samples;Finally,the γcoefficients of BN layer were used as channel sparsity factors to prune the model and remove redundant parameters to reduce the model size.Experiments were conducted under a citrus dataset in the orchard environment,to analyze the effect of network structure and Focal loss hyperparameters on the model.Compared with YOLOv4,the average precision of the YOLOv4-L detection model,constructed in this paper,was improved by 2.74%;the F1 score was 0.93,the comprehensive detection performance was better;the detection speed was improved by 2.18 times to 53.8FPS;and the model size was 15.85 MB,which was suitable for deployment in low memory devices.(2)The three-dimensional positioning model of the target was constructed.The conversion relationship between pixel coordinates of fruit and world coordinates was analyzed and derived,the color camera of RGB-D camera was calibrated to obtain the parameters required for coordinate conversion,and the depth error was analyzed to determine the best positioning range of the camera.A three-dimensional positioning model was established according to the parameters above,and the average error in the X-axis was 2.95 mm,the average error in the Y-axis was 3.45 mm,and the average error in the Z-axis was 3.67 mm measured by experiment.(3)Deployment of detection model and integrated development of detection and positioning system.To realize the application of recognition model,accelerated inference of detection models using Tensor RT enables the models to be deployed in edge devices with low computing power to achieve higher detection efficiency.Using Jetson nano as the test platform,the detection model deployment test analysis was carried out.After accelerated inference,the model detection speed was increased by2.75 times,which was 12.2 times that of YOLOv4.Using Py Qt5 integrated detection and localization system,the citrus detection and localization were completed through visual interface operation,and the functions of localization data parsing and saving were realized.(4)Kinematic analysis and motion trajectory planning of picking manipulator.The kinematics model of the picking manipulator was established by the D-H method,and the forward and inverse kinematics were analyzed.Based on the kinematic derivation,the research on the trajectory planning of the manipulator in joint space was carried out.To cope with the instability of the trajectory of polynomial planning and the need for too many constraint parameters,the trajectory planning of the quintic B-spline curve was deduced.The simulation results shown that the five times Bsample curve trajectory planning derived in this paper produces a smoother curve with fewer constraints parameters.This research provides a reliable visual detection positioning system and a theoretical basis for the robotic arm picking movement trajectory planning for the citrus picking robot. |
PDF Full Text Request