Detection and tracking of returns in sector-scan sonar image sequences

The authors report the development of algorithms for the detection and tracking of object returns in noisy sector-scan sonar image sequences. Static objects are first removed using spatial and frequency domain filtering. The optical flow of the resulting images (containing only dynamic returns) is calculated. Significant dynamic returns are detected and segmented using adaptive thresholding. The average optical flow of each significant return is used by a tracking algorithm both to constrain search window radii and to derive a similarity measure. A tree of possible tracks is maintained in which the cumulative similarity measure is used to identify the most likely tracks and to prune out the least likely object sequences. The results of experiments using real scan sequences are presented to show the utility of the proposed tracking system