Efficient fall detection is significant for the elderly, persons with motor symptoms, and people who perform risky actions, with four types of falling postures. However, most studies focused on distinguishing fall or not, although the fall direction information is crucial to turn on the corresponding part of the airbag or quickly assess the damage level. To accurately, rapidly, and reliably recognize different directional falls (forward, backward, left, and right) during daily life, this study proposes a novel fall detection methodology based on a pair of commercial lightweight smart insoles and a long short-term memory (LSTM) framework with a trained referencing denoising autoencoder (RDAE). Compared with traditional autoencoders, the referencing sub-path, i.e., RDAE, is additionally attached to achieve the automatic feature extraction. A pair of wireless in-shoe insoles (OpenGo, Moticon GmbH), each side equipped with 13 plantar pressure sensors and a tri-axial accelerometer, was employed to capture comprehensive spatial-temporal gait parameters. Hence an effective response to a fall, together with the estimation of the corresponding direction, can be accomplished, where the accuracy and response time are two primary concerns. The proposed RDAE-LSTM network provides a reliable testing result in classification, with 98.6% accuracy and 8.7 ms response time for determining fall directions, demonstrating a more convincing performance than other algorithms. The proposed methodology is an unobtrusive choice for users whose daily life is not affected by the fall detection device. The RDAE-LSTM model was proven to accurately and quickly recognize falls in four directions for the unbalanced fall detection dataset.
