A fall alert system with prior-fall activity identification
This addresses fall monitoring for the elderly by adding prior-activity identification, though it is incremental as it builds on existing detection methods.
The researchers tackled the problem of fall detection by developing a system that also identifies prior-fall activities, achieving accuracies of 88.91% for fall detection, 91.25% for fall to knees detection, and 86.25% for prior-fall activity detection.
Falling, especially in the elderly, is a critical issue to care for and surveil. There have been many studies focusing on fall detection. However, from our survey, there is still no research indicating the prior-fall activities, which we believe that they have a strong correlation with the intensity of the fall. The purpose of this research is to develop a fall alert system that also identifies prior-fall activities. First, we want to find a suitable location to attach a sensor to the body. We created multiple-spot on-body devices to collect various activity data. We used that dataset to train 5 different classification models. We selected the XGBoost classification model for detecting a prior-fall activity and the chest location for use in fall detection from a comparison of the detection accuracy. We then tested 3 existing fall detection threshold algorithms to detect fall and fall to their knees first, and selected the 3-phase threshold algorithm of Chaitep and Chawachat [3] in our system. From the experiment, we found that the fall detection accuracy is 88.91%, the fall to their knees first detection accuracy is 91.25%, and the average accuracy of detection of prior-fall activities is 86.25%. Although we use an activity dataset of young to middle-aged adults (18-49 years), we are confident that this system can be developed to monitor activities before the fall, especially in the elderly, so that caretakers can better manage the situation.