Research On Object Detection And Grasping Technology Based On Deep Learning

Robot grasping technology has great significance for service robots and industrial robots.The development of robot intelligence compels robots to the ability of autonomic learning,including grasping ability.The research of robot grasping mainly includes the analysis of the structure and function of the robot hand,the hand claw of multi finger,the qualitative analysis of robot grabbing from the form closure and the force closure,and the study of the closeness,maneuverability and stability of the robot grasping.In recent years,with the enhancement of computer computing ability,large data and machine learning have been developed rapidly,and have been applied to various fields.Deep learning has also injected new vitality into the field of robot grasping.In this paper,the target detection model and the capture network model of Robotiq140 with two finger claws are established by deep learning technology,and the grasping task of the robot is completed.The main research contents include: the influence of the virtual environment data on the training target detection network,the design of the robot grabbing parameters,the construction of the capture network mo del,and the training of the network model.Firstly,we establish a network model of target detection,and corresponding data sets to detection the household items.In order to reduce the workload of data collection,we uses the method of background separation to automatically collect data sets with labels from the virtual environment,and compares the performance of the network model trained by the data set and the real environment collected data set in different proportions.When 3000 virtual pictures and 40 real pictures were used as training sets,the average accuracy was 94%,while 320 as training set,the accuracy was 93%.Experiments show that virtual data can help network convergence.Then,the mathematical expression of grasping model is designed ba sed on the mathematical model of robot grasping and deep learning proposed by scholars at home and abroad.In this paper,we use full convolution prediction and residual unit to design the capture network model.In order to obtain the training data quickly,a vision based fixed step bounding box grabbing algorithm is proposed,and the method is used to collect training data in the virtual environment and the real environment.The collection of virtual data and real data is used to train the capture network model,and the performance of the grabbing network is verified in the virtual environment and the real environment.In real environment,the accuracy rate of grasping for familiar objects is 89%,and the accuracy of grasping is 82% in new objects(not in training objects).Experiments show that the fixed step bounding box algorithm significantly accelerates the collection of experimental data,and the use of virtual data enhances the performance of the gripping network model.Finally,a virtual simulation environment is built on V-REP,and tensorflow is used to grab robot in the virtual environment.Then the trained target detection network model and the capture network model are combined,and the real robot is used to grasp the known objects,and the success rate is 80%.Experiments show that the grasping algorithm is feasible and the whole system is stable.