• No results found

In RobWork, a robot and it’s environment can be simulated using 3D CAD models. This makes it easier to try out different scenarios and identify any faults before applying on the real robot. Thus the first step must be to draw a cad model of the setup. The conveyor belt is already installed, but unfortunately there was not any plan of the setup available. The dimensions has been measured manually and the a diagram of the setup can be seen on fig. 6.14.

The diagram of the conveyor belt was used to draw a 3D CAD model and loaded into RobWork. (see fig. 6.15, 6.16, 6.17, 6.18 and 6.19). The hook is simulated in Robwork using the pendulum model described in section 5. (see fig. 6.20).

6.5 Conclusion

In this chapter experiments were conducted to examine the precision of depth acquired from the stereo- and single view method. The rotation and the position from the robot controller were used as reference value to analyze the results. The robot-camera calibration was done to get the transformation matrix which was used during the experiments to convert between the coordinate systems. The positions of the marker in the sequence were tracked using the Kalman filter and the properties of the Kalman filter were also examined. As the last test a cad model of the scene was modeled and a simulation of the hook was done. All in all, the experiments showed that it is possible to calculate the positions with mm precision and the Kalman filter is very good at tracking these positions.

6.5. CONCLUSION 51

Figure 6.11: The plots for the nine state parameters of the Kalman filter. The red line is the reference value.

52 6 - EXPERIMENTS

Figure 6.12: The plots shows the norm of the error covariance matrix P

6.5. CONCLUSION 53

Figure 6.13: The plots for the nine state parameters of the Kalman filter. The initial velocity in y-direction is given as input. The red line is the reference value.

54 6 - EXPERIMENTS

Figure 6.14: The diagram of the setup, a) The conveyor belt; b) The hook is hang on this device; c) The cross section of the conveyor belt where ”b)” is sliding through.

Figure 6.15: The CAD model Figure 6.16: The real world

6.5. CONCLUSION 55

Figure 6.17: The CAD model Figure 6.18: The real world

Figure 6.19: The overall view of the conveyor belt

56 6 - EXPERIMENTS

Figure 6.20: A model of the hook and functions to control mass, friction, length and samplings time (Delay).

Chapter 7

Conclusion

As stated in the introduction, the most important goal of this FORK project is to examine the precision of the stereo vision. This was done in the experiments where it was shown that not only the stereo vision is very precise, but also it was better than the single view method by a factor 5.

The publicly available AR library MXRToolkit was used to find a marker which was attached to a moving robot. The algorithm is able to run in real time. The 3D position of the marker was tracked successfully using the Kalman filter and experiments were done to verify this.

A mathematical model of the conveyor belt and the hook was developed and simulated using the RobWork environment.

In this Fork, the statements stated in the introduction were tested and verified and further work in master project can be done centered around the obtained results.

In the master thesis, further development of the tracking algorithm for tracking the hook using the MXRToolkit marker and the Kalman filter will be done. Visual servoing will be done in RobWork using the CAD models of the scene and the Robot. A simulation of the hook is also available to accomplish this.

The following tasks will be examined in the master thesis:

Simulation of the placement of camera and robot in RobWork

Simulation of the Visual servoing

Tracking the hook with marker using real-time video sequences.

Marker-less tracking of the hook using model-based tracking.

The marker-less tracking will be only considered if the necessary time is available.

57

58 7 - CONCLUSION

Bibliography

[1] D. Aarno, J. Sommerfeld, D. Kragis, N. Pugeault, S. Kalkan, F. W¨org¨otter, D. Kraft, and N. Kr¨uger. Early Reactive Grasping with Second Order 3D Feature Relations. IEEE International Conference on Robotics and Automation, Rome, Italy.

[2] Martin Armstrong and Andrew Zisserman. Robust Object Tracking. University of Oxford.

[3] Jong Eun Byun and Hoon Kang. A real-time Object Tracking System Using A Particle Filter. Sungkyunkwan Univsity, Korea.

[4] Fan Xiao Charles B. Owen and Paul Middlin. What is the best fiducial? Michigan State University.

[5] Erik Cuevas1, Daniel Zaldivar1, and Raul Rojas. Kalman filter for vision tracking. Freie Universit¨at Berlin, Institut f¨ur Informatik, 2005.

[6] Tom Drummond and Roberto Cipolla. Real-Time Visual Tracking of Complex Structures.

University of Cambridge, 2001.

[7] Fakhr eddine Ababsa and Malik Mallem. Robust Camera Pose Estimation Using 2D Fiducials Tracking for Real-Time Augmented Reality Systems. Laboratoire Syst`emes Complexes, France.

[8] Franz Josef Elmer. The Pendulum Lab. 1998. http://monet.unibas.ch/ elmer/pendulum/.

[9] Richard Hartley and Andrew Zisserman. Multiple view Geometry in computer vision.

Cambridge university press, second edition, 2003.

[10] Michael Isard and Andrew Blake. CONDENSATION – conditional density propagation for visual tracking. University of Oxford, 1998.

[11] Anders Kjær-Nielsen, Dirk Kraft, and Lars Baunegaard With Jensen. Using high-resolution cameras for detection of open-/closedness of cylindrical objects. The Mærsk Mc-Kinney Møller Institute - University of Southern Denmark, 2007.

[12] Vincent Lepetit and Pascal Fua. Monocular Model-Based 3D Tracking of Rigid Objects:

A Survey. Computer Vision Laboratory, Lausanne Switzerland, 2005.

[13] Vincent Lepetit, Francesc Moreno, and Pascal Fua. Selected Topics in Computer Vision.

http://cvlab.epfl.ch/teaching/topics/.

59

60 BIBLIOGRAPHY

[14] N. Pugeault, S. Kalkan, E. Baseski, F. W¨org¨otter, and N. Kr¨uger. Reconstruction un-certainty and 3D relations. Proceedings of Int. Conf. on Computer Vision Theory and Applications (VISAPP’08), 2008.

[15] Luca Vacchetti, Vincent Lepetit, and Pascal Fua. Fusing Online and Offline Information for Stable 3D Tracking in Real-time. Swiss Federal Institute of Technology.

[16] Greg Welch and Gary Bishop. An Introduction to the Kalman Filter. University of North

Carolina at Chapel Hill, 2006.

Timetable for the Fork

Figure 1: Timetable for the Fork

As with any other project, there have been some delays and setbacks. In the beginning the idea was to record image sequences in different directions to compare each other. During the fork it was found that the rectification of the images was not right, but the undistortion worked fine. Therefore the images were only undistorted and the missing rectification didn’t affect the experiments. Then the sequence which was used in the experiments was recorded and in the end of December the rest of the sequences in the other directions captured. However it was discovered that right camera has be moved in the period between the recording of the first sequence and the other ones. This meant that the projection matrix was altered and the depth calculation was wrong. Fortunately the first recorded sequence was correct and this was used in the experiments.

61

Related documents