Wavelet thresholding-based multi-layer perceptron for fall detection

Fall detection and waveform identification pose substantial challenges in healthy monitoring of humans. In this paper, we propose a novel method, namely wavelet thresholding multi-layer perceptron, which harnesses the power of multi-layer perceptron applied to threshold-processed wavelet features, for simultaneous fall detection and waveform identification. The proposed method outperforms in detection accuracy and model interpretation. Wavelet transform concurrently extracts time-frequency information, while thresholding effectively compresses data via sparse representation. Simultaneously, the powerful ability of learning nonlinear correlation enables our method to acquire the relation between fall and gait data more effectively. We further reconstruct trajectories associated with falls and daily activities using selected important features. This approach amplifies the interpretability of differences between fall and non-fall trajectories. The new method is applied to two public datasets to illustrate the significance of analyzing gait characteristics and exploring applications in monitoring falls. A comparison with other methods demonstrates the superiority of our method.