3D Object Modelling

ALT="Persistent Vision Pty Ltd"

3D object modelling from rangefinder data

ALT="Puma robot arms"

While with the Intelligent Robotics Research Centre at Monash University, John contributed to a robot-vision project, using a coded-stripe rangefinder developed by Rick Alexander. The objective was to demonstrate modelling of a polyhedral object using multiple views, allowing the object to be manipulated and oriented by a robot arm and gripper. The rangefinder was mounted on a second robot arm. All computations were performed in C on a Silicon Graphics workstation.

This work was supported by a GIRD grant and by BHP as the industrial partner.

Technical Summary

Each light-stripe from the rangefinder created a plane intersection with the object being measured, allowing any intersected edge to be located as a 3-dimensional point lying within the plane. Using multiple stripes, a visible edge of the polyhedron was represented by a set of 3D points, nominally lying on a straight line, but subject to some error.

Line Fit Diagram

A single rangefinder view of the object provided a set of 3D points, whose association with edges had to be inferred. The fitting of straight lines to subsets of points, and the grouping of points into edge-subsets, required an iterative process to resolve the ambiguous points near polyhedral vertices. A fairly complex statistical technique was used to obtain lines of best fit during this process.

After determining the 3D lines representing edges visible within a view, the vertices were located from the edge-intersections. Of course, the lines would not intersect exactly in 3D space, so a statistical fit was used to estimate the positions of vertices and the edges joining adjacent vertices.

One rangefinder observation only provided information about the visible parts of the object. By making observations from six directions, most polyhedral objects could be observed in their entirety. The last phase of the process was to merge the results from six individual views, creating a database model of the object.

One application of the model would be to use it to determine the orientation of a similar object picked up by the robot gripper. The new object would not necessarily have to be observed from more than one direction.

John's Contributions

The method used to locate an edge-point within a light-stripe cross-section, and the derivation of error estimates for the edge-point co-ordinates.
The statistical analysis used to obtain a line of best fit in three dimensions, accounting for the arbitrary three-dimensional error-estimate of each point.
The iterative process used to resolve which points belonged to each line.
Design and coding of the algorithms used for analysing rangefinder data and deriving lines, vertices and a merged wire-frame model of the test object.
Designing and constructing an electrically-rotated transparent table, with accurate set-points 90 degrees apart, to orient the object within the rangefinder's field of view.

Skills

C programming
3D co-ordinate geometry
Statistical analysis of errors
Turning and fitting

Reference

J.M. Badcock and R.A. Jarvis, "Wire-frame modelling of polyhedral objects from rangefinder data", Robotica Vol 12, pp 65-75, 1994.

Back to experience summary page
Back to home page