Visiir

Vision based track search for a RoboTour2006-like robotic competition.

Goal

The goal of this project is to find a track in front of the robot. This task is crucial not only for the track following, but also for the junctions recognition.

Not a goal

One can hardly expect a 100% reliable simple method for a natural scene recognition any time soon. And it is also not a goal of this project. Any "where is the track?" hint is better than going blindly.

Features of the algorithm

• Unsupervised learning: No need for a huge training dataset.
• Online learning: It is able to learn a new track type when approached.
• Extremely limited set of "magic" parameters: only k and n for a k-means algorithm (part of the whole algorithm) with a fixed amount of n clustering cycles. The n parameter can be omitted when a different termination condition is used. The k is suggested to be 3.
• Robust: As shown bellow, the same algorithm works well on images from different cameras taken in a different time under different light conditions.

Algorithm description

Not here yet. It would be nice to use it for the RoboTour 2007 Delivery Challenge in October. Thus it makes no sense to make it public in advance.

Video

1_25fps.avi (20.3MB, XviD avi, 1min 36s) shows an on-the-scene track recognition. With the progress the task becomes gradually more and more difficult (water reflections, narrow track, mud ...). 1_12fps.avi (41.2MB, XviD avi, 3m16s) shows the same in slow motion. Better for a more carefull visual inspection.

2_25fps.avi (6.5MB, MSMPEG4, 1min 4s) shows how the algorithm deals with a highly structured environment. The travel starts on a concrete pavement and switches over an asphalt road to a very difficult mud path.

3_25fps.avi (25.8MB, MSMPEG4, 4min 28s) deals with many road material changes - asphalt, concrete, sand.

4_25fps.avi (19.6MB, MSMPEG4, 3min 21s) test the algorithm robustness in rainy conditions. The black path with many extremely light reflections hides between very dark mud and darkly brown bushes in winter.

Illustrational images

	Scene	Visiir track recognition (darker in areas with a higher likelihood of being on track)	Vissir using different features (slower processing)
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.