Fall detection using doppler radar and classifier fusion

Liang Liu, Mihail Popescu, Marilyn Rantz, Marjorie Skubic

Abstract: Falling is a common health problem for elders. It is reported that more than one third of seniors 65 and older fall each year in the United States. We develop a dual Doppler radar system for fall detection. The radar system generates a specific Doppler signature for each human activity which is then categorized by a set of classifiers as fall or non-fall. However, different classifiers may produce different decisions for the same signature. In this paper, we propose a fusion methodology based on the Choquet integral that combines partial decision information from each sensor and each classifier to form a final fall/non-fall decision. We employ Mel-frequency cepstral coefficients (MFCC) to represent the Doppler signatures of various human activities such as walking, bending down, and falling. Then we use three different classifiers, kNN, SVM and Bayes, to detect falls based on the extracted MFCC features. Each partial decision from a classifier is represented as a confidence. We apply our fusion method to a dataset that consists in 450 activity samples (109 falls and 341 non-falls).