Autonomous on-wrist acceleration-based fall detection systems: unsolved challenges

Fall detection (FD) has been the focus of many research studies during the last years. Developing reliable FD systems is relevant, for instance, to provide support to the elderly population in their everyday life. Besides, the generalization of the use of wearable devices (and more specifically, on-wrist devices) to measure the daily activity strongly suggests that in a short period of time, the elderly people will be making use of this type of devices. On-wrist devices can be used as the FD basic sensing unit; while the intelligent classification can be obtained either autonomously (on the device) or requested to a remote service (via the paired smartphone or via web services). This study tries to analyze the current challenges in autonomous on-wrist wearable devices for producing a reliable and robust FD system. To do so, we analyze the related work; one of the possible solutions is implemented with several alternatives and evaluated with publicly available simulated falls data sets. The most remarkable findings in this research are that i) real fall data sets are needed, at least, a valid merging method to produce real fall like Time Series, ii) generalized solutions might not be enough and research is needed in models that learns from the user, iii) the need of tuning and fitting to the current user performance, iv) the amount of fall types suggests that hybrid and ensemble approaches might be interesting. (c) 2020 Elsevier B.V. All rights reserved.

Automated summary

Research on fall detection systems is crucial for supporting the elderly population in their daily lives. Challenges include the need for real fall data sets, studying models that learn from users, adapting to user performance, and exploring hybrid and ensemble approaches.