Adaptation of a partial shape recognition approach

Abstract

This paper presents the adaptation of a partial shape recognition method and an application to underwater sonar images. To recognize partially occluded objects, each one is represented by a set of "landmarks". The landmarks of an object are points of interest that have important shape attributes. Objects contour changes according to the sonar settings and the acquisition conditions. Thus, the location of the objects points of interest is not preserved. To take into account these disturbances, three major adaptation procedures are implemented: 1) a more robust landmarks extraction method is proposed; 2) the classical approach is adapted to handle multiple occurrences of objects in the scene; and 3) the initial least square match error is replaced by a Fourier descriptors calculation. Thus, the match quality evaluation is based on matched contour portions instead of matched landmarks. Last step of the partial shape recognition is the decision strategy that consists in thresholding the match error. This adaptation of the method can be generalized to other kinds of application, like robotic or computer vision.