It is very important to accurately inspect machining errors, product assembly tolerances in the manufacturing industry. Typical three-dimensional measurement methods include a coordinate measuring machine (CMM), a line laser scanning method, and a structured light system that includes a camera and light source for generating a pattern. In general, the inspection system applying the three-dimensional measurement method requires physical calibration processing using special equipment to place the object in the home position with the desired pose.
Therefore, to solve this problem, this thesis proposed a method for measuring randomly placed objects by coordinate recognition. It is assumed that the position and position of the object is varied with each measurement. Coordinate of CAD model must be brought to coordinate of measured data to calculate deviation of object.
The first step is rough registration based on principal component analysis and the nearest point iterative algorithm. The second step is the main methodology of this thesis, which is to adjust the coordinates to calibrate the transformation parameters. Coordinate adjustment consists of two stages, which are reference plane matching for the calibration of rotation parameters and edge matching for translation parameters.
Introduction
- Background
- Motivation
- Objective
- Outline of the thesis
At this time, there is not enough time to approximate the position of the product or adjust the origin. Therefore, it is necessary to develop a technique to automatically recognize the coordinates of randomly positioned objects. The product in process is positioned randomly as it moves on the conveyor or when the robot does not place the product in a predetermined location.
Even if the position and attitude of the measuring object is different from the ideal one, a system capable of recognizing the coordinate and performing the inspection is needed (Figure I-1 (a)). In addition, the coordinate recognition algorithm has a field to be used besides the measurement system. Coordinate of CAD model must be brought to the coordinate of measured data to calculate deviation of object (Newman et al., 1995).
The method should even deal with an object that is randomly placed with the variation of X, Y, and Z translation along the X, Y, and Z directions and the rotation angle around the X, Y, and Z axes. The details of the proposed 3D registration algorithm for randomly placed object measurement are described in Chapter 3.
Literature Survey
- Contact
- Non-contact
- Target based matching
- Feature based matching
- Point based matching
This method uses cameras and processes the images to reconstruct the 3D surface of the object (Cui et al., 2013). Therefore, Son et al. propose an automated measurement system process; the system consists of a laser scanning device and a software module. Therefore, the accuracy of the ToF device depends on how precisely the time is measured (Gokturk et al., 2004).
Therefore, correct object detection is also part of a natural user interface (Cui et al., 2010). Traditionally, objects should be placed in the starting position with special fasteners or tools that are completely dedicated to certain products (Y. Li et al., 2006). Mathematically, registration means finding the optimal transformation matrix between DCS and MCS (Abenhaim et al., 2011).
Shi & Liang used uncoded reference points to identify 3D coordinates of product and align point cloud sets with a common coordinate system such as reference CAD model (Shi et al., 2016). There are different types of features, usually including basic geometric shape such as lines, polygons or arcs (Laboureux et al., 2001). The main advantage of this method is lower computational cost of the fitting phase (Dutagaci et al., 2012).
This approach requires well-structured elements of the environment, consisting of a large flat surface and regular geometric shapes (Konecny et al., 2016). Castellani et al., 2008) The high salience of the mesh indicates that the corresponding point is important for different scales. In general, the threshold is set to the average of the total visibility of the grid over the local maxima.
FPFH is based on a histogram of the difference of angles between the normal of the neighboring points (Rusu et al., 2009) It is an advanced version of Point Feature Histogram (PFH) that retains the discriminating power of the PFH. A transformation parameter for both rotation and translation is calculated between two different data (Dai et al., 2007). To start, the ICP algorithm needs initial guess to iteratively change it (Rusinkiewicz et al., 2002).
3D Deviation Analysis of Randomly Positioned Objects
Problem Statement
The target application of this approach is inspection of the assembly quality of the products in the process and analysis of the deformations of the finishing of the car body. Also, the entire inspection process must be fully automated and run in real time so that an inspection can be carried out during the process. Localization, so-called registration, is the process of transforming point cloud scan data from measurement coordinate systems (MCS) to design coordinate system (DCS) (Mehrad et al., 2013).
The usual way to register algorithm is to perform a coarse registration, which roughly aligns the two data sets in the same coordinate, and then perform a fine registration (Figure III-2) (Bellekens et al., 2014). Most of the research on registration uses a new method of fine registration, Iterative Closest Points (ICP). Although ICP makes it possible to search for a more optimal solution than the result obtained through coarse registration, but it often reaches a wrong convergence.
In this paper, the technique of adjusting coordinates for optimal alignment after performing fine registration, which is ICP, is presented. This system uses only the cloud of points immediately captured by a 3D measuring device for real-time application. In general, the registration algorithm is also used to merge point cloud datasets that are collected from different angles.
However, scan-to-scan matching, which is point cloud registration from multiple cameras, is not covered in this paper. The purpose of this study is to accurately match the scanned data to the CAD model for deviation analysis.
Registration and Adjustment Processes
- Overview
- Preprocessing: Reference plane selection
- Rough registration
- Coordinate adjustment
This is necessary for matching the proposed technique to the reference plane in the coordinate matching method, which is the third step. In conventional registration processes, ICP is performed after feature-based matching, but in this thesis, ICP must be preceded to find a reference plane. However, we cannot directly detect which point set corresponds to the reference plane at the time the data is scanned.
In the scan data, the points closest to the reference plane on the CAD model are estimated as reference planes. Then it is possible to perform coordinate adjustment about the rotation parameter by reference plane adjustment. This is preliminary work using a method called reference plane matching to be applied to the coordinate alignment step. The user must in advance select a suitable face as a reference plane on the web of the given CAD model.
Also, after that, the measured data is registered in the corresponding proximity level, and many points are likely to be registered in the reference level. Therefore, it is recommended to select the reference plane by assigning priority to the face of the mesh in order of width. We will perform reference level matching sequentially on aircraft that most closely meet the above conditions.
First, a reference plane matching method is introduced to correct the rotation matrix that determines the shift angles of the x, y, and z axes. Reference plane matching is a method of matching two planes, which are a reference plane selected from CAD mesh and the corresponding plane extracted from measured data. Point sets recorded on the reference plane would be used in the next step. a) Aligned CAD and scan data (b) Point sets registered on the reference plane.
Now we can compare the normal vectors between the CAD reference plane (𝑛⃗⃗𝐶𝐴𝐷) and the approximate plane from the scan data (𝑛⃗⃗𝑠𝑐𝑎𝑛). Matching a reference plane using the difference between two normal vectors Repeat steps 1 through 3 when multiple reference planes are selected. As can be seen in Figure III-19, the data obtained by performing the rough registration and reference plane matching are well aligned in x, y, and z displacement angle, but the translation seems to need some adjustment.
The difference between the centroids of two data was used to derive the translation vector while performing reference plane matching, which seems insufficient for accurate translation matching. First, the reference plane matching is performed to project the point cloud data and the scan data from the CAD that fit the XY plane accurately to some extent.
Summary
Translation Parameter is derived directly by comparing centroid of initial point set with transformed point set.
Experiments
Experiment Setup
Experiment Results
- Variation in the shift angle
- Machining error detection
With the proposed algorithm, it is shown that the cumulative error value drops to less than 1˚ overall. The average accumulated error of the ICP was 2.0245 degrees and it is almost 5 times the average value of the proposed method 0.4246. As shown in Figure IV-7, the 3D deviation plot of the proposed registration (a) shows that the whole area of the normal product is blue, which means that it is not defective.
However, the 3D deviation plot applied to the ICP (b) looks like about 3 mm defects appear at the end of the product, even if the tested part is a normal product. Therefore, if the registration result is not accurate, it may cause a wrong result to identify the fault. The average recorded time was 12.468 seconds, the highest recorded time was 14.502 seconds, and the lowest recorded time was 10.869 seconds.
It is also set to be 150 mm away from the design coordinate system in the x-axis direction and 60 mm away from the design coordinate system in the y-axis direction. Deviation graph clearly shows the area of defect on the surface of object with red color.
Conclusion and Future Research
Conclusion
Future research
Paper presented at AMBIENT 2014: Fourth International Conference on Ambient Computing, Applications, Services and Technologies, August Rome, Italy. Paper presented at the 28th International Symposium on Automation and Robotics in Construction, Seoul, Korea. Accuracy assessment of iterative closest point registration for different phantom surfaces captured by an optical 3D sensor in radiotherapy.
Fixtureless profile inspection of non-rigid parts using the numerical inspection fixture with improved definition of shear boundary conditions. Paper presented at Proceedings of the 2nd int conf on machine control guidance, Bonn.
