The task of recognizing objects from ultrasonic echoes using an adaptive and mobile sonar sensor is described. The biomimetic sensor consists of a center transmitter flanked by two receivers that adaptively rotate to maximize the echo bandwidth.The sonar is located at the end of a robot arm and moves in response to the echo time of flights to position an object along the transmitter axis at a known range and elevation to maximize the incident acoustic energy. A learning stage is first employed,followed by a recognition stage. The learning stage uses ultrasonic echoes from each object at its various poses with respect to the sonar and selects informative features from the echoes to form a data base. The recognition stage processes an observedecho to extract the informative feature vector and forms a classification based on the minimum distance of an observed point to a database entry. The sonar differentiates between the head and tail side of a coin in air (λ =0.6cm).This paper describes improvements in the sonar that makes it more robust and efficient. The time-of-flight is determined in a more robust manner by initially detecting a large echo and then searching backward in time to find the threshold crossing point.The data window size is made adaptive to capture only significant echo waveform data. The feature vector then has a variable size, providing structure to the data base to make the search more efficient. Separate feature vectors are determined for theright and left receivers rather than concatenating the two to form a single feature vector. This allows the object recognition to be more robust in the presence of masking echoes.
展开▼
