Fall Detection Using Time2Vec and Transformers
| dc.contributor.advisor | Ngu, Anne H. H. | |
| dc.contributor.author | Noorani, Shahroz | |
| dc.contributor.committeeMember | Metsis, Vangelis | |
| dc.contributor.committeeMember | Gao, Byron | |
| dc.date.accessioned | 2023-12-19T21:20:41Z | |
| dc.date.available | 2023-12-19T21:20:41Z | |
| dc.date.issued | 2023-12 | |
| dc.description.abstract | Falls in older adults can have many lasting health and financial problems if not handled properly or swiftly. The WHO estimates that, on average, 35% of older adults over the age of 60 face a fall during the year, and this number only increases with older age groups (World Health Organization, 2008). Addressing this critical concern, this research ventures into augmenting fall detection mechanisms by harnessing the capabilities of Time2Vec alongside Transformer models. Our goal is to aim for accurate detection of falls so that needed help can be rendered as soon as possible. Drawing upon established machine learning paradigms, we performed extensive experiments employing transformer encoder layers on different data types and sensors. Time2Vec was incorporated as a vectorization layer to address the limitation of transformers in dealing with sequencing in the absence of a positional encoder. Examining various feature subsets, we established that vectorizing the x-axis of the accelerometer data using Time2Vec significantly enhances the stacked encoder model’s performance, creating a robust mechanism for fall detection. This adjustment led to an optimized model capable of efficiently identifying falls, thereby holding substantial promise in mitigating the adverse impacts associated with falls in the elderly population. Our endeavors resulted in an overall improvement of 12% in F1 scores, especially in larger datasets compared to previous architectures. The findings from this research contribute to the ongoing efforts in enhancing fall detection and underscore the potential of Time2Vec and Transformer models for monitoring other health conditions when time series data are involved. | |
| dc.description.department | Computer Science | |
| dc.format | Text | |
| dc.format.extent | 57 pages | |
| dc.format.medium | 1 file (.pdf) | |
| dc.identifier.citation | Noorani, S. (2023). Fall detection using Time2Vec and transformers (Unpublished thesis). Texas State University, San Marcos, Texas. | |
| dc.identifier.uri | https://hdl.handle.net/10877/17796 | |
| dc.language.iso | en | |
| dc.subject | fall detection | |
| dc.subject | Time2Vec | |
| dc.subject | transformers | |
| dc.title | Fall Detection Using Time2Vec and Transformers | |
| dc.type | Thesis | |
| thesis.degree.department | Computer Science | |
| thesis.degree.discipline | Computer Science | |
| thesis.degree.grantor | Texas State University | |
| thesis.degree.level | Masters | |
| thesis.degree.name | Master of Science |