Vehicle track detection in CCD imagery via conditional random field

Coherent change detection (CCD) can indicate subtle scene changes in synthetic aperture radar (SAR) imagery, such as vehicle tracks. Automatic track detection in SAR CCD is difficult due to various sources of low coherence other than the track activity we wish to detect. Existing methods require user cues or explicit modeling of track structure, which limit algorithms' ability to find tracks that do not fit the model. In this paper, we present a track detection approach based on a pixel-level labeling of the image via a conditional random field classifier, with features based on radial derivatives of local Radon transforms. Our approach requires no modeling of track characteristics and no user input, other than a training phase for the unary cost of the conditional random field. Experiments show that our method can successfully detect both parallel and single tracks in SAR CCD as well as correctly declare when no tracks are present.