| dc.description.abstract | Large numbers of false clutter detections arise in the use of low frequency active sonar for the detection of low Doppler submarine targets in shallow water. Traditional detection algorithms operate individually on each beam output searching for targets at all ranges. Detection algorithms such as the Page test are designed for target echoes that are extended in range owing to the multipath propagation of shallow water channels and reflection off of the target. However, the target echo and bottom features may extend over multiple beams either physically or by bleeding through the sidelobes of the beamformer. This indicates that detections need to be associated across bearing or the detector must be designed to account for targets and clutter that are spread over multiple beams. This report considers the detection/association issue from an image processing perspective, applying a Markov random field (MRF) model to the image formed from the range-bearing sonar data. The Markov random field model exploits a priori information such as the distribution of the reverberation, minimal information about the distribution of the target echo and the relationship between each rangebearing cell and its neighbouring cells. Maximum a posteriori (MAP) estimates of the labeling of each range-bearing cell (i.e., target or reverberation) are obtained rapidly through an iterative algorithm with an initialization provided by the Page test detector output. The objects in the resulting range-bearing image are then tested according to their size, ruling out any too large to be submarines. Application to real data has shown a reduction of greater than 80% from the number of Page test detections over all the beams to the number of submarine-like objects in the MRF-MAP image. | |
