The factors that make stereo change detection effective can have deleterious effects on automatic terrain extraction algorithms. This fact can be exploited for change detection. Areas of poor correlation (due to image differences) get a low Figure of Merit (FOM). The terrain posts within these areas can be color-coded to indicate areas where one suspects change has occurred. The same suggestion was pursued to develop NGATE.

The theory of stereo image matching is that conjugate pixels are from the same fixed ground object. NGATE performs stereo image matching on every pixel. With sufficiently high resolution and accurately registered stereo images, users can match most pixels even on comparatively patternless surfaces such as highways. To compute the elevation of a pixel, a user multiplies a constant by the amount of X parallax. The amount of X parallax is the shift of a conjugate pixel between stereo image pairs.

When a vehicle moves, the conjugate pixels are no longer from the same location. If the stereo image matcher is able to match pixels from moving vehicles in two scenes, the amount of X parallax is wrong and results in an elevation blunder. Depending on how much a vehicle moves between two images, the stereo image matcher will make either "no match" or a "false match" in the area near the vehicles. In the no-match case, the vehicle has moved too far for the stereo image matcher to find matching pixels in both frames. In the false-match case, the vehicle is correlated in both frames, but has exaggerated X parallax due to displacement. An elevation blunder results. Over fast-moving highways, moving vehicles generally result in no-matches. Over slow-moving traffic, false matches are common. There are also other phenomena that can produce false- and no-match results (see "Minimizing Undesirable Effects" below). 
 

As part of NGATE enhancements, BAE Systems developed an algorithm to automatically detect and remove elevation blunders caused by moving vehicles. NGATE is able to judge the legitimacy of pixel matches, using rule-based logic. Suspect elevations get a low FOM. During terrain extraction, NGATE then will apply a special filter to eliminate implausible spikes. In addition to being an automatic terrain extraction algorithm, NGATE is, in effect, an automatic moving vehicle detector.

What's on the Drawing Board?
The idea of using stereo image matching technology to detect change, including moving vehicles, is worth further investigation. BAE Systems' research indicates this idea may lead to technology that can successfully detect moving vehicles in accurately registered stereo images of suitable resolution. Stationary vehicles can also be detected using a combination of bare earth algorithms and stereo image matching algorithms inside NGATE. More work must be done to reduce false alarms, because change and motion aren't the only reasons no-match or false-match results occur.

Combining algorithms that operate on radiometric properties and image geometric properties may lead to further success in change detection and vehicle detection. NGATE doesn't currently attempt to measure the velocity vector of moving vehicles, so this should be an area of future research. Nor does NGATE have the ability to definitively classify objects as vehicles. For now, it simply recognizes that a vehicle-sized object has caused correlation to fail and repairs the DEM in that vicinity.

As part of an alerting system, NGATE would send the polygonal boundaries of the presumed moving vehicles to analysts for assessment. In an orchestrated workflow, the analyst's image exploitation application would automatically load the imagery at each candidate target location. The analyst would make the final judgment as to whether the object is, in fact, a vehicle, and record the information accordingly. NGATE processes images quickly and could prove valuable as an automated change detection algorithm.
 

  See  Featured Images
 
  Subscribe to Earth Imaging Journal

  Go to Home Page