Effect of data representations on deep learning in fall detection

Fall-related injuries can have a significant impact on the quality of life of the elderly population. Because of the upward trend in the elderly for continued independent living, there is a growing need for reliable fall detectors that can enable prompt assistance in case of falls. Doppler radar technology offers a number of desirable attributes for realization of fall detection and health monitoring systems that can facilitate self-dependent living. Human motions generate changes in Doppler frequencies that can be accurately captured using time-frequency representations. A variety of time-frequency distributions have been proposed in the literature. In this paper, we investigate the impact of different time-frequency representations on the performance of a deep neural network based fall detector. Using real data, we demonstrate that the choice of data representation in the time-frequency domain is important for enhancing the accuracy of the fall detector.