Visual System Design And Grasping Simulation Of Strawberry Harvesting Robot Based On Edge Depth Neural Network

The strawberry is a vital cash crop,but its planting and harvesting process entails significant manual input.Automating its harvesting can reduce labor costs and increase economic benefits.However,the complex strawberry planting environment and the small size of the object pose challenges for automatic and accurate recognition and intelligent picking.As a result,achieving these goals under the guidance of machine vision algorithms has become a popular research area.The progress of deep learning has significantly enhanced the ability of visual detection algorithms to identify objects in complex environments.Nonetheless,deploying these algorithms on embedded devices to achieve fast and accurate detection has received limited attention.This article aims to design and accelerate a deep neural network for strawberry detection deployed in edge embedded systems.Simulation experiments on picking robotic arms guided by the above detection methods will also be conducted.The article focuses on the following topics:(1)Aiming at the problem that the deep neural network on edge equipment has high precision and low speed in strawberry detection,which can’t meet the needs of automatic picking,this paper improves the YOLO network and designs a new lightweight network,which significantly improves the detection speed while ensuring the accuracy of strawberry detection.The experimental steps mainly include two aspects.Firstly,the original CSPNet structure of YOLOv4-tiny is replaced by CSPNet＿F and CSPNet＿S.Secondly,the backbone network is redesigned,and four new lightweight networks are designed.After synthesizing the indicators of accuracy and speed,this paper chooses RTSDNet network as the strawberry detection model.The model can successfully detect strawberries in complex situations such as shadow coverage,leaf occlusion and close aggregation,and effectively improve the detection speed,which provides the possibility for the model to be deployed on edge embedded devices.(2)The accuracy of deep neural network is higher than that of traditional image processing algorithms,but the model of deep neural network is large,which makes it more difficult to deploy.The deployment of deep neural networks can be divided into two ways,cloud deployment and edge deployment.Cloud deployment has high network requirements,high energy consumption,and large delay caused by network transmission.It is difficult to meet the requirements of strawberry picking in the field.Therefore,edge deployment is selected in this paper,and Jetson Nano is used as the edge deployment device.In order to further improve the detection speed of the model at the edge,the experiment quantifies the model through the Tensor RT method.The experimental results show that the quantization method can nearly double the detection speed,and the accuracy of the model remains basically unchanged.It shows that the method of Tensor RT quantization model is effective and feasible,and the real-time detection of the model on embedded devices is realized.(3)The present study employs the Gazebo software to create a simulation environment for autonomous grasping of a robotic arm in the ROS system.To test the effectiveness of the system,obstacle avoidance experiments are conducted on 2D and3 D maps.The RRTconnect algorithm is chosen as the path planning algorithm for the robotic arm,and Moveit is used to control its movements.The study further incorporates the quantified RTSDNet model of Tensor RT with the robotic arm grasping system to conduct simulation grasping experiments on multiple strawberries and a single strawberry.Results demonstrate the feasibility and reliability of the proposed path planning algorithm.