Design Of Mushrooms Grabbing Device With Binocular Vision Manipulator Based On Deep Learning

With the continuous improvement of mushroom production,the modern mushroom food processing industry has put forward increasingly high requirements for the automation level of the entire mushroom industry.After removing the handle of mushroom,the preparation time of the production line before food processing can be greatly shortened,automatic production is imminent.In order to solve the problem of capturing mushroom before removing its handle in food processing line,a robot arm device based on deep learning and binocular vision is designed to capture mushroom.The main research contents are as follows:Firstly,according to the literature review and research,the current development status of mushroom industry,target detection and industrial arm is introduced and analyzed,and the background and significance of this device is clarified.On the basis of this,the function requirements and the overall design scheme of the device are clarified.According to the requirements,a binocular visual mechanical arm grabbing mushroom device based on deep learning and embedded platform is proposed.The device is divided into target detection system,binocular stereo positioning system,and arm grabbing system as a whole.Secondly,this paper complete the construction of target detection system.By comparing and analyzing with mainstream object detection algorithms,and according to the specific needs of the mushroom grabbing device,YOLOv5 was selected as the target detection algorithm.The mushroom detection dataset was made and expanded by ourselves,and various training strategies were compared and the best scheme was selected.Thirdly,this paper complete the construction of binocular stereo positioning system.Complete the pixel coordinate system.Based on the monocular camera imaging model class,a binocular camera imaging model is developed,and the binocular camera stereo imaging model is built.The measurement error is reduced by distortion correction,polar constraint and stereo correction.Finally,the specific internal and external parameters of binocular camera are calibrated using Zhang Zhengyou calibration method.Then,this paper complete the construction of the mechanical arm grabbing system.The hardware selection of each part of the three-degree-of-freedom manipulator is completed,and the physical construction of the manipulator capture system is completed.Set up a physical prototype of the three-degree-of-freedom mechanical arm system.By deriving the forward and reverse kinematics formulas of the arm,the rotation speed of each step motor and the rotation angle of each joint are calculated when the position of the end executing mechanism of the arm is known.the track generation and path planning during the movement of the arm are completed.Finally,this paper complete the deployment and experiment of each subsystem.The deployment and communication series of each subsystem were completed on the embedded platform Jetson Nano.In the process of target recognition,Tensor RT engine was constructed for accelerated reasoning and Deep Stream was used to preprocess video images.The performance indexes of each subsystem were determined.The whole feasibility of the device is verified.The experiment shows that the device can realize the real-time recognition of target mushroom stably and quickly,and grasp the target mushroom.It has good applicability and robustness,and it can meet the requirements of design and use.