|
|
Trevor Taylor - Research InterestsI am interested in a couple of quite different fields: Computer VisionI am finishing a PhD part-time in the area of Computer Vision for autonomous robots. Despite over 30 years of research in this field, there are still many open questions. This is a dangerous statement to make in the presence of computer vision researchers! However, anyone who is familiar with the field will realise that we still have a long way to go to match the Human Visual System. My work was broken into three phases:
HistoryYou can read a copy of my original Stage 2 Proposal document from 2002 that has been converted to a web page. However, it is now quite out of date and there have been several changes. My Confirmation document is available in PDF format. This provides an updated overview of my PhD plans and progress at that stage.
At the time of my Confirmation I demonstrated a Hemisson robot navigating around obstacles
in a "playing field" (referred to as "Tupperworld" for reasons that will be obvious
if you view the videos) in a purely reactive manner using vision from a Swann wireless camera.
The Hemisson had the unfortunate drawback of an umbilical
cable, which is quite apparent in the videos.
(For those who are interested, the top-down views were taken using a Sony DSC-92 Digital Still Camera, which is why they are MPEG files. The view from the Hemisson is a collation of all of the BMP files saved during a test run. I use BMPs because they are not lossy like JPEG, which is important when I want to play back the test run later and have it re-processed by the program. This video was created under Windows XP using Movie Maker, which only produces WMV files. The ASF file type also works with Media Player, and seems to be the same as WMV.) I switched to a Yujin soccer robot which can be controlled via a radio modem so it does not have the umbilical problem. However, it is quite small and cannot move around freely on carpet. This limits its usefulness, but it has been a good platform for early experiments. To build maps, I introduced the concept of "Pirouettes" -- the robot spins around on the spot taking a series of snapshots with the onboard camera. Because I was using cheap cameras, the robot has to stop and "settle" after each small rotation before taking a picture. I typically use 12 rotations of 30 degrees each, which gives a 50% overlap between successive images. Working with hardware is a pain. So I developed a simple simulation using OpenGL that was based on the Yujin robot. This allowed me to work without being limited by the robot hardware and the small size of the "playing field" that I had built. I later developed a replacement for this very simplistic simulation using the much more sophisticated simulator in the Microsoft Robotics Studio. The simulation I built (based on code oroginally from Ben Axelrod) is known as the Maze Simulator. It has been used by people from around the world based on the feedback I have received. The simulated robot could map its environment using an Occupancy Grid. It used a Distance Transform to find paths to the nearest unknown space and so it could tell when there was nothing left to explore. I even did some work with multiple collaborating robots. In 2005 I purchased an X80 robot from Dr. Robot in Canada. (I bought another one in 2006, which in retrospect was not a smart move.) X80s have on-board pan and tilt color cameras, WiFi networking, and many other sensors. An X80 is large enough to roam the corridors of an office block. Although the robot was supplied with software, it was an ActiveX control which was not appropriate for my application. I therefore developed the necessary code to incorporate it into my own program. I have written an interface for the robot in C++ that allows me to teleoperate an X80, and another version in C# that will run on a PDA with a WiFi interface. This software is open source. See my X80 WiRobot Page for downloads. I used an X80 to explore and map areas of the office building where I was working. The biggest problem turned out to be the lighting. Although we humans can easily navigate the corridors, they are poorly lit. The camera on the X80 does not work well in low light, so the robot sometimes loses sight of the walls. In 2006, Microsoft released Microsoft Robotics Studio. This has proven to be a major distraction from my PhD, and I have spent a lot of time working with MSRS. Some of my work is on my MSRS Code Page. This eventually culminated in co-authoring a book on MSRS which is due for release in mid-2008. A major setback occurred when I tried to implement SLAM and discovered that the input data from the vision system is not sufficiently rich to guarantee convergence of a particle filter. The problem is very easy to understand once you realise what is happening. My vision system has a field of view of 60 degrees. Also there is a maximum range of only a few meters before the uncertainty in range measurements reaches the point where the data is too unreliable to be used. This is a consequence of perspective.
Contrast this with a Laser Range Finder which is
the most popular device for SLAM research. A LRF has a 180 degree
field of view and is accurate to about 1cm out to a range of
20 meters, or even 100 meters for a top-quality LRF.
To visualize this, the
diagram below shows the area seen by a LRF as the large
semi-circle, and the reliable vision area as the small
pie slices in the center which show different vision
ranges based on increasing the camera resolution. Can you see why LRFs work and vision does not? Orders of magnitude more data, and far more accurate too! To overcome this problem, I made the assumption that walls are always at right angles and run "North-South" or "East-West" in the map. Although this allowed me to build reasonable maps, my supervisory committee were not happy with this restriction. One of my tasks is therefore to try to relax this limitation. At present I am finalising my PhD. I have completed my Final Seminar, but the review panel requested major changes to my Thesis, in particular related to the way that I had structured the document and the style I had adopted which was unconventional and more suited to a textbook. I still need make these changes. I also want to migrate my code to the MSRS environment. Computer Vision Publications (in reverse chronological order)
Teaching & Learning in IT EducationWeb RepositoryThe Web Repository, or WebRep for short, is the subject of a small Teaching & Learning grant within the Faculty of Information Technology at QUT. The objective is to provide an area where students can upload web sites that they have developed as part of units undertaken at QUT. WebRep fulfills several objectives:
The WebRep was deployed in Semester 1, 2005. Evaluation was undertaken by surveying students and there was an overwhelmingly positive response. Students like to be able to see the work of their peers and to be able to show off their own work. There are now hundreds of web sites in the WebRep and it has been used for six semesters.
The following paper discusses the use of the Web Repository in teaching
Web Development at QUT:
TALSS Teaching FellowshipIn 2005 I was a Teaching and Learning Support Services (TALSS) Teaching Fellow (Scheme B -- one day per week). My project for the year consisted of developing a branching mechanism for use in the QUT OLT (Online Learning and Teaching) system. This will allow academics to embed questions in OLT pages to lead students through various paths. The primary objective of this project was to combine my existing online tutorials and quizzes to produce an environment I refer to as Online Interactive Learning (OIL). The expectation was that this would engage the students and make the tutorials more interesting. The requirement to answer questions along the way should increase the level of concentration necessary to complete the tutorials. Unfortunately, it was not possible to reach my ultimate goal, i.e. to use existing quizzes and record the answers. However, I have introduced multi-way branching into OLT. Although this concept is not really new, it is a facility that was not previously available in the OLT system. In 2006 QUT undertook a major evaluation of Learning Managment Systems and I was involved on the Academic Review Panel. After selecting Blackboard, QUT migrated from OLT to Blackboard in 2007. Although I am still interested in developing interactive tutorials, the changes to the Learning Management System meant that I did have not pursued this line of research further. [Home Page | Robotics] |
