Human fall detection methodologies: from machine learning using acted data to fall modelling using myoskeletal simulation

Human Fall Detection is a research area with interest from many disciplines and aims to perform for many assisted-living monitoring applications to promptly identify life-threatening situations. A fall occurs when a person is unable to maintain balance due to a variety of issues; physical; mental or environmental. The accurate detection of the fall is crucial as a missed detection can be fatal. Variability of human physiological characteristics is currently unstudied as to the impact on a fall detector's performance as young adults and elderly are expected to fall differently. Another important issue is the scene occlusions. In the use of visual sensors, an occluded fall is treated as a missed detection as the whereabouts of the person is unknown when occluded. Finally, current studies are based on acted fall datasets on which algorithms are trained. These dataset are unrepresentative of real fall events and illustrate the events without occlusions or other scene in uences. Several fall detection algorithms were developed during the study aiming to achieve accuracy in detection falls while fall-like actions such as lying down remain undetected. Human fall datasets were used for training and testing purposes of A machine learning algorithm using data from depth cameras which captured the fall events from different views. A new pathway was introduced tackling the issues of availability issues of data-driven machine learning approaches which was achieved with the use of simulation data. The use of myoskeletal simulation was then selected as a closer representation of the human body in terms of structure and behaviour. With the use of a simulation model, a personalised estimation of the fall event can be achieved as it is parametrised on a physical characteristic such as the height of the falling person. Alternative technologies such as accelerometers have been used for fall detection to prove the validity of this approach on other modalities. A study regarding the impact of occlusions for fall detection which is one of the issues not properly investigated in current work is proposed and examined. Synthetic occlusions were added to existing depth data from publicly available datasets. The research methodologies were evaluated using the most representative depth video and accelerometer data from existing datasets, as well as YouTube videos of real-fall events. The machine learning methodologies achieved good results on similar body variability datasets. A discussion regarding the proof of concept of the simulation-based approach for fall modelling is mentioned given the comparative results against existing methodologies which achieves better than any existing work evaluated against known datasets. The simulation approach is also evaluated against occluded fall and non-fall event data, proving the further robustness of the approach. This platform can be expanded to analyse any type of fall, or body posture (e.g. elderly), without the use of humans to performs fall events.