Comparison of Human Fall Acceleration Signals Among Different Datasets

Falls can have a major impact on physical and psychological health on elderly people who experience them. To reduce the negative consequences of a fall event, automatic fall detection systems are being developed to correctly identify when a person falls and alert the caregivers to provide assistance on time. Performance of fall detection algorithms is tested with datasets containing measurements of falls and regular activities of daily living. In this work we acquired fall signals with accelerometer sensors and compared them with digitized fall signal records from 6 different datasets. Additionally, three threshold-based algorithms for fall detection were implemented and their performance was tested with the analyzed datasets. The results suggest that a heterogeneity among the fall data in distinct datasets exist and that it affects the performance measures of tested datasets.