2019.11.33
Before applying
Kalman filtering, the false positives for the case of falls were
60% of the total classified instances, and after applying Kalman
filtering were minimized to 33%, respectively. For annotation
purposes, the three movement types were associated with three
integers: 1 for walk; 2 for run; and 3 for fall, respectively. Actual
run and fall events are also annotated on the diagrams. For each
experiment, two different diagrams were generated: one illustrating classification results based exclusively on acceleration
data and one illustrating classification based on both acceleration and sound data. As it is indicated, Kalman filtering improves
the overall detection by smoothing the sequential occurrences
of run or fall events, respectively. In addition, the use of sound
as additional classification feature has increased the accuracy of
fall detections by minimizing the false ones in cases of simple
walk and of walk and run. A threshold t = 10 has been selected
for determining the occurrence of fall or run events from the
total sequence of classified movement types (i.e., if sequential
occurrence of fall movement types >10, then a fall is detected).