Experiments in the piece-wise linear approximation of ultrasonic echoes for object recognition in manipulation tasks

This paper describes a novel method of object recognition based upon the piece-wise linear approximation of ultrasonic echoes which, when tested with examples typical of a work cell environment, achieves classification success rates of 92-98%. The approach uses decision tree classifiers constructed from simple features derived from the echoes of a monostatic sonar and taken from a number of view points. The results illustrate the effect upon the method's classification success of the use of a single view point, the inclusion of additional view points, and the possibility of object rotations, for 22 classes of objects. The processes of feature extraction and classification are accomplished within a time comparable with the time of flight of the pulse, and hence the method has potential for real time applications.