Lines of research

THE VARIOUS FALL DETECTION MODES:LINES OF RESEARCH*
  (*source: N. Noury, TIMC-IMAG Laboratory, University of Grenoble I)

Detecting changes in the rate of vertical motion

This detection system is based on identifying fast acceleration in the rate of vertical motion, something that occurs during a fall

As it rises linearly over time, the vertical speed that corresponds to the constant acceleration due to the gravitational pull will exceed the speed threshold of controlled motion during a fall.

The Wu team at the University of Vermont in the USA has shown, by analyzing video footage of patients falling, that the vertical and horizontal speeds were dissimilar during controlled motion but became practically similar during a fall. Hence the modifications to these speeds is a good way to characterize a fall. Nait-Charif and Rougier too studied falling by analyzing video footage, but in this case through head motion and using the vertical and longitudinal speed thresholds.

The main problem with these detection methods is that the detection thresholds are hard to determine due to the significant variability in the controlled motion speed of individuals (the variability between individuals). If these thresholds are too low, they lead to the detection of falls that are not falls at all (false positives). On the other hand, if these thresholds are too high, then there is the risk of missing undetected falls (false negatives). A number of teams are attempting to overcome these error causes (Noury (University of Grenoble) through supervised learning or Depeursinge (CSEM, Lausanne) through unsupervised learning).

A number of different kinds of sensors have been put forward, especially accelerometers and gyroscopes (rotation), with a localization at the upper body level (Hwang, Noury), in line with the sacrum (Prado, Spain), at the waist (Matie) or even using the head.

To go beyond just simple detection, a Japanese team (Fuyaka, Japan) recently proposed that critical phase detection could trigger the deployment of an airbag!


Detecting the impact with the ground

On impact, the motion speed is suddenly reversed. This detection system is based on identifying the reversal in the direction of acceleration.

The sensors used are accelerometers and impact detectors, with a localization at the waist level (Williams and Doughty, Ireland), at the center of gravity (Lindeman, Germany) or at the upper body level (Bourke, Ireland).

The main difficulties arise when it comes to determining the direction of the trajectory along with the "signature" of the impact, the cause of false positives.


Laying down on the floor

These detection systems seek to identify when a body is horizontal, meaning that it has fallen.

There are a number of techniques available: Some sensors distributed over the body detect a horizontal position ("dead body"), or a loss of contact between feet and the ground (Tamura, Japan). However, when the person lies down on a bed or on a sofa, the detectors may take that as a fall (false positive).
Another method used is sensors in the floor that detect the body's contact with the floor ("actimetric floor").

It should be noted that some laboratories are interested in the noise of impact at the time of the fall: they are developing noise sensors that can isolate the impact noise and raise the alert. Naturally these techniques must be perfectly specific so as to avoid raising an alarm every time a book is dropped or a pair of glasses!


Prolonged immobilization

Detection is based on a lack of motion

The sensors that detect the immobility that is supposed to follow a fall may use various methods: accelerometer vibration meters (when all motion stops), infrared presence sensors or even video cameras.

In this way, the motion of different body segments has been analyzed using a mobile phone (Tamura, Japan). The absence of any motion has also been analyzed by a ceiling mounted video camera (Mihailidis, University of Toronto) or using infrared presence sensors (Noury, University of Grenoble).

If the immobilization time between two motions that are considered to be critical by the device, prior to raising the alarm is too short, then false positives are obtained. If on the other hand, this idle time is too long, then the reaction time will be too long.

Video raises a number of technical problems in terms of control over the field of vision, ambient light or vision in space. This method also gives rise to ethical problems due to the intrusion that it represents.
One proposal is to integrate decision making algorithms directly at the camera level so as to avoid transferring the picture outside of the home.


Scenario analysis

Detection is based on a combination of parameters that are characteristic of a critical situation

Various sensors record data linked to the environment conditions prevalent at the time of the fall and what happens immediately afterwards, information that is then evaluated by algorithmic analysis.

Measures have been developed from cross referencing actimetric data recorded by infrared sensors, microphones or even video cameras equipping the home.

Other mechanisms that are especially promising are based on an algorithmic analysis of cross referenced data from accelerometers that detect falling and infrared detectors that detect prolonged immobilization (Noury, University of Grenoble).