Feature-based Object Recognition Method For Robotic Grasping

Robot grabbing is the concentrated expression of intelligent robot system,and target recognition,as the core step of robot grabbing,has always been a research hotspot in robot field.With the continuous improvement of computer computing power and sensor imaging quality,the detection and grabbing of target recognition based on simple two-dimensional image in structured environment has been widely and deeply studied,and relevant technologies have gradually become mature.However,with the increasing demand for intelligent robots,only two-dimensional information is used to represent three-dimensional objects in complex environment,which inevitably leads to information loss,making it difficult for robots to achieve high-precision grasping of multi-objective objects in unstructured environment.Thanks to the continuous development of point cloud acquisition technology and computing power in recent years,the research on point cloud vision and grasping has become a new hotspot in the field of robotics.However,object recognition and robot capture based on point cloud is still a challenging and valuable problem.In view of this,this paper proposes a high-performance point cloud local descriptor-point to point feature histogram(hoppf)from the perspective of 3D point cloud,aiming at the problem of model and scene representation in capture,and develops a compact and efficient binary descriptor b-hppf based on this.Then,aiming at the problem of target recognition and location,two kinds of three-dimensional target recognition algorithms based on the above descriptors are proposed,and on this basis,the research on robot grasping of multi-target objects is carried out.More specifically,this paper has made the following progress in point cloud local feature extraction,3D target recognition,robot grabbing system calibration and multi-target recognition grabbing:1.In the aspect of point cloud local feature extraction,firstly,a point to point feature histogram(hoppf)descriptor is proposed,which has superior comprehensive performance in discrimination,robustness and calculation efficiency.The AUC value of the corresponding RPC curve can reach 0.9070,which is 0.5105 higher than the experimental base-line.Secondly,in order to meet the strict memory requirements and matching rate of the robot end Based on hoppf,a binary point to feature histogram(b-hppf)descriptor is proposed to further improve the performance of the descriptor in feature matching efficiency and structure compactness.The compactness parameter of the descriptor is 3.75,which is 15 times higher than the classical real value descriptor and 3 times higher than the other binary descriptors.The extraction time in the local neighborhood is 0.1056 seconds,which is 15 times faster than the other descriptors.2.In the aspect of three-dimensional target recognition,based on the descriptors hoppf and b-hppf,two three-dimensional target recognition algorithms are developed: one is based on random sampling consistency,the other is based on rotating centroid clustering.Then we compare the performance of target recognition accuracy and attitude estimation accuracy with the classical descriptors on each authoritative data set.The corresponding target recognition score,the accuracy can reach 92.25%,the rotation error is 0.5168%,and the translation error is 1.152 mr,which verifies the effectiveness and practicability of the proposed algorithm.3.In the aspect of robot grabbing system calibration,this paper proposes a calibration method of robot dual arm base coordinate system,the accuracy of which is 0.5210 mm after calibration;a pixel plane hand eye calibration method and a point cloud space hand eye calibration method,the accuracy of which is 1.075 mm;the accurate calibration results provide a reliable mathematical basis for robot grabbing.Then,based on the above algorithm,a calibration software is developed,which integrates hand eye calibration and double arm calibration,completes the data visualization of the calibration process,and makes the calibration link of camera and robot simple,reliable,stable and easy to operate.4.In the aspect of robot multi-target recognition and grabbing,firstly,the proposed hoppf and b-hppf descriptors and three-dimensional target recognition algorithm are used to realize multi-target recognition and pose estimation in point cloud space;secondly,based on simple two-dimensional image information,an efficient stack object detection algorithm of rotating point set projection and a centroid calibration algorithm of fitting projective coordinates are developed to calculate The average running time of the method is 0.102 seconds,and the success rate of stack detection can reach 100%.Then,the robot grabbing research of multi-target objects is carried out based on the above algorithm.The experimental results show that the grabbing success rates of single target objects,non stack objects and stacked objects are 96.7%,97.8% and 90%,respectively.