Research On Unmanned Logistics Sorting Method Based On 3D Vision

With the development of the times,due to people's desire for a better life and society's requirements for production efficiency,the use of robots instead of humans for simple repetitive labor has become the trend of the times.Robots can free humans from harsh or dangerous work environments,increase work efficiency and reduce production costs.In the field of logistics sorting,it is mainly based on manual sorting.Due to the low efficiency of manual sorting,when the peak of the logistics is unable to cope,and the manpower operation is not standardized,it causes many unnecessary losses.Therefore,the machine replaces the manual in this field.Very necessary.However,robots need to have a pair of human-like eyes that can perceive rich information,such as the color,type,and size of objects.Traditional RGB cameras can only acquire image color information,but two-dimensional imaging can no longer meet the robot's perception requirements.The task of the robot is becoming more and more complex,so the robot's perception of the three-dimensional world is essential.With the development of sensors,TOF cameras provide a simpler and more intuitive depth perception method that allows the camera to have the ability to sense the three-dimensional world.In this paper,TOF camera is applied in the field of logistics sorting.The method of collaborative complementation between TOF camera and RGB camera is proposed.The convolutional neural network classification algorithm based on deep learning is adopted to make the sorting equipment have the function of intelligent classification and measurement.Compared to manual sorting,this method can digitize the results of the classification measurements and upload information for optimal decision making.This paper mainly completes three tasks: one is to complete the camera calibration research,the other is to complete the image classification algorithm research,and the third is to complete the dimension measurement algorithm research.The main research contents are as follows:(1)The principle of camera imaging is studied,and the mathematical model of the camera with or without distortion is studied.The principle of Zhang's calibration method is derived.The calibration work was completed for the two cameras,and a depth measurement calibration scheme was designed for the TOF camera.(2)Based on the convolutional neural network,a transfer learning method is used to train an RGB classification neural network with an accuracy of 90% in the case of a small data set.(3)A dual-input single-output network was designed,using RGB images and 3D depth images as inputs to two input layers.The classification result is combined with the color information and the depth information to make up for the blind spots of the two kinds of image information,and the classification accuracy reaches 83.2%.(4)Based on the 3D depth image and the RGB image,a mathematical model for measuring the volume and size of the object is established,and the dimensional measurement error is less than 2 cm.The classification accuracy and measurement accuracy of the unmanned logistics method studied in this paper are basically applicable to the logistics sorting system with different shape,color and volume of the target object,and the To F algorithm is simple and real-time.