AUTOMATIC INTERPRETATION OF SONAR IMAGERY USING QUALITATIVE FEATURE MATCHING

This paper, reports recent work on the automatic interpretation of sector scan sonar imagery. Previous work has resulted in a blackboard-based system, employing a mixture of image-, signal-, and rule-based processing to extract appropriate feature information from sonar scans. We here report on the continuation of this work, describing a system capable of carrying out classifications of observed objects based on available feature measures, such as size, shape, and gray level characteristics. The problem of determining feature measures which are invariant to changes in sonar setting, object position/orientation, and noise characteristics is addressed by using qualitative measures to describe object features during matching for recognition. Typical attributes are ''large,'' ''small,'' or ''highly variant.'' Invariance comes from dynamically selecting the threshold values used to map the numerical feature values derived from the image data to these qualitative measures. Descriptions of the qualitative appearance of known objects are maintained by the system as ''exemplars.'' Recognition therefore takes place by matching observed object descriptions to exemplars in either a constrained or unconstrained fashion. Descriptions are presented for the feature measures used, the quantitative-to-qualitative mapping, and the matching procedures, with results showing the discrimination provided by the feature measures, the changing numerical boundary values between qualitative attributes, and the overall success of the recognition processing for single sonar scans. The overall interpretation is shown to be 86% successful for objects viewed on different sonars in different conditions, provided features measures giving good discrimination between objects are employed.