Automatic interpretation of sonar image sequences using temporal feature measures

This paper reports the development of a system for the automated interpretation of sector scan sonar data. It proposes the use of a new combination of feature measures derived from sequences of sonar scans to characterize the behavior of targets' returns over time. Previous research used grey-scale and shape descriptors derived from single sonar scans. However, problems were experienced with targets whose return varied significantly over time (such as divers, UUV's, and ships' wakes). Hence a new set of temporal feature measures has been developed by combining existing one-dimensional temporal measures and two-dimensional object descriptors. These new features provide a quantitative description of the behavior of a target's two-dimensional returns over a sequence of sonar scans. Experiments with a limited but real data set have shown that classification accuracy can be significantly improved by the use of these new features. The use of "static" feature measures (derived from a single scan) was observed to give classification errors of between 7% and 10% when they were applied to the data set. In contrast, the use of temporal measures reduced this error rale to 1% or 2% and in some cases reduced it to zero.