Crop-picking Robot Grasping System Based on Deep Learning

The crop-picking system plays a key role in automatic picking. In this study, a set of crop-picking robot grasping system based on convolutional neural network (CNN) was designed by using deep learning technology. Firstly, a variety of preprocessing methods are used to expand and process the image data to ensure the diversity and integrity of the sample data. Then, CNN and gradient descent (GD) algorithm were combined to optimize the training model to identify the types of crops accurately. Then, the lidar used ToF (Time of Flight) technology and AGglomerative NESting(AGNES) algorithm to obtain the key positioning points. After excluding invalid cluster centers, the identification results were accurately located. The experimental results show that compared with the 3D pose estimation method of the manipulator based on multi-view, the calculation error of the 3D coordinates of crops is smaller, and the reasoning speed for a single image is only 0.000316 seconds. Finally, the comprehensive accuracy of the system reaches 97.58%, which provides practical technical support for automatic crop picking.