Falling detection concerns identifying and alerting when a person falls or experiences a sudden loss of balance. It is an important safety feature, particularly for elderly or vulnerable individuals, as falls can lead to severe injuries. One applicable technique is employing Long Short-Term Memory (LSTM) networks that can be used for falling detection by processing data from sensors, such as accelerometers or gyroscopes, to identify patterns associated with falls. This MatLab source code demonstrates that a single LSTM layer can be used for this purpose. Still, multiple LSTM layers can improve the system's performance with less processing time, reaching 99.13% of accuracy with 4.35% of loss values. Furthermore, this work thoroughly studies the sensitivity of the size of an LSTM cell, demonstrating that it is not necessary to obtain a size close to the timestamp of the features and thus save hardware resources.
Cite: J. P. Matos-Carvalho, S. D. Correia, S. Tomic, "Sensitivity Analysis of LSTM Networks for Fall Detection Wearable Sensors", 2023 6th Conference on Cloud and Internet of Things (CIoT'23), Lisbon, Portugal, 2023, pp. 112-118, doi: 10.1109/CIoT57267.2023.10084906