Back
to indexT. Boult and A. Erkin and P. Lewis and R. Michaels and C. Power and C. Qian and W. Yin
Summary
This paper presents a high-level description of techniques for identifying moving objects, such as humans, from omni-directional cameras. An omni-directional camera produces a true 360 degree FOV image of the scene around the camera. The authors note that "unwrapping" the image to produce tracking over the resultant perspective image would be accurate, at a considerable time expendature. One of the goals of the program was to find the objects quickly, when they are far away and therefore consist of fewer pixels. An interesting point is made, where the spatial resolution along the horizon is greatest, compared to the spatial resolution of nearby objects, because of the curvature of the mirrored lens used to produce the omni-directional image.
Methods
The Lehigh Omni-directional Tracking System (LOTS) starts with a background image, and uses a change-detection algorithm to isolate pixels that have changed frame-to-frame. To account for small movements common in the frames, such as leaves of trees waving around, a second background image is stored and compared against as well. Target objects are generally more than one pixel wide, so a clustering algorithm is used to find adjacent pixels that move together. Over sucessive frames, these pixels move slowly throughout the image, and are therefore labelled as a possible target. Many possible targets can be selected from this technique.
I don't think this will work from a moving vehicle, since its designed for a static image.
Keywords
omni-directional camera, motion tracking, surveillance.
Rating
7
Bibtex Entry
@article { boult98frame,
author = "T. Boult and A. Erkin and P. Lewis and R. Michaels and C. Power and C. Qian and W. Yin",
title = "Frame-rate multi-body tracking for surveillance",
journal = "Proc. of the DARPA IUW",
year = "1998",
url = "citeseer.nj.nec.com/177759.html"
}