Acoustic fall detection using one-class classifiers
Mihail Popescu, Abhishek Mahnot · 2009
Falling represents a major health concern for the elderly. To address this concern we proposed in a previous paper an acoustic fall detection system, FADE, composed of a microphone array and a motion detector. FADE may help the elderly living alone by alerting a caregiver as soon as a fall is detected. A crucial component of FADE is the classification software that labels an event as a fall or part of the daily routine based on its sound signature. A major challenge in the design of the classifier is that it is almost impossible to obtain realistic fall sound signatures for training purposes. To address this problem we investigate a type of classifier, one-class classifier, that requires only examples from one class (i.e., non-fall sounds) for training. In our experiments we used three one-class (OC) classifiers: nearest neighbor (OCNN), SVM (OCSVM) and Gaussian mixture (OCGM). We compared the results of OC to the regular (two-class) classifiers on two datasets.