Object Classification from Acoustic Analysis of Impact

Abstract

We address the problem of autonomously classifying objects from the sounds they make when struck, and present results from different attempts to classify various items. Previous work has shown that object classification is possible based on features derived from the frequency content of signals. We develop a moving-maximum algorithm to extract the two most significant spikes in the FFT of the sounds of impact, and use these extracted spikes as features. We describe the transformation of the training data? extracted features into a compilation of representative cluster means. These cluster means are used as labeled inputs to the different classifiers discussed. We discuss two techniques to classify test vectors based on their extracted feature spikes, and show that accurate classification of objects is possible using these features. The first technique is the familiar minimum-distance classifier that calculates the distance between a given test vector and each cluster mean, and assigns the test vector to the cluster that yields the smallest error. The second technique is one we developed for the task: the decision-map classifier, a hybrid minimum-distance classifier and decision-free classifier, that iteratively finds the closest cluster mean to each test vector and uses multiple features only if it cannot classify the test sample. Results from classifier trials show that using our moving-maximum feature extractor and decision-map classifier, object classification from the sound of impact can be done as accurately as using the minimum-distance classifier, but at significantly lower computational expense.