Research On 3D Point Cloud Based Positional Measurement Method For Curved Frame Parts

The production and processing of curved frame parts is an important part of the automotive production automation line.Currently,most of the loading and unloading systems for curved frame parts in automotive production automation lines are still an open-loop system.The system uses positioning pins or blocks to limit the position and attitude of the workpiece,and then corrects the gripping by a fixed amount of compensation measured manually,and finally the robot arm completes the gripping.Since the system uses positioning pins or blocks and is open-loop,the robotic arm cannot correct the gripping position in real time and the gripping accuracy is not high.Therefore,it is necessary to introduce a vision system to make the loading and unloading system a closed-loop system to achieve accurate loading and unloading.In order to solve the above problems,this thesis adopts Kinect V2 camera as the data acquisition device,and conducts a research on the 3D visual position measurement method for curved frame parts,so as to achieve the requirements of running rate t<3s,position error Δd<4mm,and angle error Δθ<2° for the visual position measurement method of curved frame parts in actual working conditions.The main research contents of this thesis are as follows.（1）According to the actual working conditions,a suitable visual guidance system is selected as a model to design the subsequent posture measurement method,the relationship between the coordinate systems in camera calibration and hand-eye calibration and the realization principle are studied,and the camera calibration is completed using the Zhang Zhengyou checkerboard grid calibration algorithm.The Tsai-Lenz hand-eye calibration algorithm was used to complete the hand-eye calibration,and the accuracy of the hand-eye calibration was verified,with the rotation angle error of 0.5° and the average position offset error of 0.4mm to meet the actual working conditions.（2）For the problem of cavity noise in the depth map collected by Kinect V2 camera,the cavity repair algorithm is studied,using bilateral filtering to repair small cavity noise,and proposing a cavity repair algorithm based on the combination of RGB map and depth map,using the edge information of RGB map to assist the depth map to repair large cavity noise.Then the conversion from depth map to point cloud data is realized and the conversion accuracy error is verified to be 0.6 mm,which can meet the actual working condition.（3）The point cloud segmentation and recognition algorithm of curved frame parts is studied.A point cloud segmentation algorithm based on the template plane is proposed for the actual working condition,and the segmentation of workpiece and table or conveyor is completed by using the mode of making the template plane when offline and segmenting according to the flatness and position relationship between the template plane and the real-time plane when online.The Euclidean clustering segmentation algorithm is used to segment between workpieces,and the number of points in the point cloud and the volume of the point cloud are used as criteria to identify curved frame parts according to the actual working conditions.Finally,the overall average operation rate of the point cloud segmentation and recognition algorithm for curved frame parts is 0.186 s,which can meet the actual working conditions.（4）The point cloud alignment algorithm for curved frame parts is studied.An improved Super-4PCS algorithm combining ISS-3D features is designed for the characteristics of curved frame parts with rich geometric features.The improved operation rate is verified to be better than the original algorithm.Finally,it is verified that the average running speed of the proposed improved Super-4PCS+ICP algorithm is 2.425 s,the average error of the two axes is 2.435 mm,and the average error of the rotation angle is 1.153°,which can meet the requirements of the actual working conditions.The final total solution has an average run rate of 2.611 s,an average error of 2.835 mm in both axes,and an average error of 1.653° in the angle,which can meet the requirements of the actual working conditions with the run rate t<3s,position error Δd<4mm,and angle errorΔθ<2° in the attitude measurement.