Estimation of the "best angle for next view" in a sequential multiple aspect classification scheme
Résumé
Classification of three-dimensional targets using two-dimensional sensor information is a standard problem in sonar processing. Improvements of the classification performances are generally reached by improving the sensor's resolution and/or by considering multiple views of a target. Enhancement of the sensor's resolution considering single view information can improve the ability of a specific classifier to discriminate against different targets. However, if the target is located in a complex environment, its signature can be altered by a phenomenon of occlusion and this method does not necessarily lead to better classification performances. The use of multiple perspectives offers an interesting alternative to solve this problem. The main objective of this study is to increase the rate of correct classification, and to minimize the classification time, using a fusion of multi-view information by replacing the current strategy (for which the number and orientation of the views are defined a priori) by an adaptive approach. The discussed adaptive approach is based on the result of a primary single view classification step and uses the already acquired information to either stop the classification task if the classification result is unambiguous or define the best direction for the next view (the most discriminative) if the classification remains ambiguous. Classification steps are performed using a measure of similarity between the value of a computed (extracted) features vector and a set of different model-issued possibilities. The developed algorithm has been tested on a large set of simulated data (simulation of 900kHz REMUS-100 images). The results were then compared to a single view and multi-aspect classifier that uses a predefined angular (azimuth) step. Benefit