Research On Object Grasping And Recognition In Warehouse Environment

In recent years,with the rapid development of the e-commerce industry,the sorting efficiency of the logistics warehousing system directly affects the time of product distribution and the performance of the company.The current logistics and warehousing system mainly relies on manual sorting to select goods.Due to the rising labor costs in recent years,the logistics company's transportation volume has increased year by year,resulting in the company's operating costs becoming larger and larger.In order to solve this problem,the establishment of high-efficiency,high-quality automated warehousing systems is becoming a research hotspot for related companies and researchers.Automated sorting is the core problem of automated warehousing system research.In order to study this problem,this project builds a robotic grasping and recognition automation system for warehousing environment based on ROS,PCL and TensorFlow.This project uses a suction and a two-finger jaw as gripping actuator.A gripper switching device for the gripper of the suction and the two-finger jaw is designed.It allows the suction and two-finger jaw to be used for different objects while performing the grasp task,or to collaborate on the grasp task when a single gripper cannot complete the grasp task.In the method based on the suction,in order to obtain the target grasp point,the point cloud clustering and model selection algorithm are used to segment the point cloud on the surface of the object,and the centroid point of the plane is calculated as the grasp point.For the problem that the model selection algorithm sometimes has error segmentation,a moving least squares(MLS)optimized model selection algorithm is proposed.The point cloud is preprocessed by moving least squares method,and the point cloud with smaller Gaussian curvature fluctuation and smoother surface is obtained.Then the problem of attribution of each surface slice is determined by minimum description length criterion(MDL).In the jaw-based gripping method,the parameters of GQ-CNN were modified for the experimental hardware device as a jaw-based grasping algorithm,and the gasping success rate of GQ-CNN was tested in the experiment.In order to improve the grasping success rate of the gripping system based on the suction and the jaw,this paper constructs the best grasping strategy.The experiment shows that the strategy has a significant effect on the success rate of the grasping.For the automatic identification of objects,the shallow convolutional neural network is established considering that the weight parameters of the existing convolutional neural network are too more and too difficult to train.The data set is built to train the network.In order to improve the recognition robustness of the network,the data set is expanded ten times using data enhancement.The recognition accuracy of the network on the new data set is 87.5%.The experimental platform was set up to test the grasping and recognition system based on the suction,the grasping and recognition system based on the jaw and the grasping and recognition system based on both actuators.The quantitative analysis method was used to summarize the performance of the system.At the same time,for the work that the suction and the jaw can not be completed separately,the experiment of the cooperative operation of the suction and the jaw is designed,and the importance of the gripper switching device is verified.