Performance of Optical Flow Techniques
for Continuous Motion Mobile Robot Navigation

Chris McCarthy

Abstract | Thesis download | Video | Image Sequence | Publications

Thesis Abstract

This thesis presents a new methodology for the in-system comparison of vision techniques for mobile robot navigation. We apply this methodology to a class of vision techniques for estimating optical flow (visual motion). Previous comparisons of optical flow and associated techniques have compared performance in terms of accuracy and/or efficiency, and only in isolation. These comparisons are inadequate for addressing applicability to continuous, real-time operation as part of a robot control loop. In this thesis, we argue that to accurately assess suitability of vision techniques to robot navigation tasks, comparisons must consider the performance of the technique as part of a system, and not just in isolation. In addition, performance must be gauged on criteria that reflect the control needs of the system. We argue this in the specific context of optical flow techniques. The methodology prescribes two sets of comparisons: off-board and on-board. Each is conducted in the context of specific, but representative tasks for identified classes of vision-based mobile robot navigation. In so doing, different perspectives on the performance of vision techniques are provided through the combination of off-board and on-board results. For the comparison of optical flow techniques, we assess performance in the context of two visual control scenarios: corridor centring and visual odometry. Both represent distinct control needs for mobile robot navigation in corridor-like environments. Four optical flow techniques, including one correlation-based method, and three gradient-based methods, are included in this comparison. In addition, three spatio-temporal pre-filters are considered in combination with the gradient-based methods examined. Techniques are chosen on the basis of their distinct design properties, considered representative of a large proportion of existing optical flow methods. The comparison results form the basis of a set of recommendations of optical flow techniques for mobile robot navigation. The comparison results also provide insight into the core issues that govern the in-system performance of optical flow techniques, and vision algorithms in general, when embedded in a control loop. Through the application of this comparison methodology, more general outcomes concerning the use of any vision technique in closed-loop control are also provided. It is demonstrated that in-system performance of vision techniques can vary significantly from performance in isolation. Despite the presence of flaws/constraints in a vision technique, the algorithm may still perform well in the system, provided that the system controls its speed and direction to keep within the parameters of the algorithm. Further to this, it is shown that the required timeliness of the vision algorithm response is dictated by the needs of the task. If the robot controls its speed to its environment, slow executing vision techniques may still perform well. This thesis also demonstrates that the accuracy of vision techniques is not as important as suitability to task. The underlying vision algorithm response must satisfy the control needs of the system, thus emphasising the need for such in-system comparisons.

Thesis Download

The thesis can be downloaded as a PDF document:

Masters Thesis Video Repository

The following two videos have been included as supplements to the results presented from on-board trials conducted for the thesis. They provide a sample of relative performances achieved by various optical flow techniques compared in this thesis. None of these particular trials were used to obtain results presented in the thesis. Please note there is some overlap of content in these video's.

All video's are in WMV format.

Side Wall Image Sequence

Below is the full side wall image sequence used for offboard comparisons. Images are in pgm and ppm format.

Publications

  1. "Comparison of temporal filters for optical flow estimation in continuous mobile robot navigation", by C. McCarthy and N. Barnes. Proceedings of ISER'04 International Symposium on Experimental Robotics, Singapore, 2004 (in press).

  2. "Performance of optical flow techniques for indoor navigation with a mob ile robot", by C. McCarthy and N. Barnes. Proceedings of IEEE International Conference on Robotics and Automation (ICRA2004), pp 5093-5098, New Orleans, US A, 2004.

  3. "Performance of temporal filters for optical flow estimation in mobile r obot corridor centring and visual odometry", by C. McCarthy and N. Barnes. Proceedings of the 2003 Australasian Conference on Robotics and Automation. Brisbane, Australia, 2003.


Created by Chris McCarthy.
Dept. of Computer Science and Software Engineering,
The University of Melbourne.
Last Updated: 23 Feb, 2005.