Use of kinematic and mel-cepstrum-related features for fall detection based on data from infrared depth sensors

Abstract A methodology for acquisition and preprocessing of measurement data from infrared depth sensors, when applied for fall detection, combined with several approaches to the classification of those data, is proposed. Data processing is initiated with extraction of the silhouette from the depth image and estimation of the coordinates of the center of that silhouette. Next, two groups of features to be applied for a fall/non-fall classification are extracted: kinematic features (various statistics defined on the position, velocity and acceleration trajectories of the monitored person) and mel-cepstrum-related features (components of the mel-cepstrum obtained by means of an unconventional set of mel-filters). Finally, the utility of these features in fall detection is assessed using three classification algorithms − viz. support vector machine, artificial neural network, and naive Bayes classifier − trained and tested on two datasets consisting of, respectively, 160 data sequences (representative of 80 falls and 80 other human behaviours) and 264 data sequences (representative of 132 falls and 132 other human behaviours). The application of the combination of the kinematic and mel-cepstrum-related features yields highly accurate classification results − all classifiers achieved, depending on the dataset, 98.6–100% and 93.9–97.7% sensitivity. Thus, infrared depth sensors can be promising tools for unobtrusive fall detection. They provide data which can be in various ways preprocessed to form a basis for reliable fall detection. Appropriate selection of the feature sets directly affects the reliability of unobtrusive monitoring systems, and − indirectly − the quality of life of the monitored persons.