Publicação
Markov transition field for fall detection using time-series data
| Resumo: | Fall detection systems have traditionally relied on sequential pattern recognition methods, using, for example, time series data obtained from inertial sensors, such as accelerometers. This paper proposes a methodology for fall detection based on converting time series from accelerometer sensors into visual representations using the Markov Transition Field (MTF) method. The UP-Fall dataset was used to test the performance of a Convolutional Neural Network (CNN) model trained on the MTF images generated. A systematic analysis of the image generation parameters was carried out, including the window size, the percentage of overlap, and the number of bins used in the discretizations. The experiments showed that the configuration with 55 bins, a window of 200 samples, and 40% overlap resulted in the best accuracy (97.13%), demonstrating that the conversion of sensory signals into MTF images is a promising alternative for fall detection, allowing computer vision models to capture relevant temporal patterns with high efficiency. |
|---|---|
| Autores principais: | Kalbermatter, Rebeca B. |
| Outros Autores: | Silva, Felipe G.; Pereira, Ana I.; Valente, António; Lima, José; Yahiaoui, Réda; Fayad, Moustafa |
| Assunto: | Fall detection Markov transition field Convolutional neural network Time series segmentation Sensor data |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | comunicação em conferência |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Instituto Politécnico de Bragança |
| Idioma: | inglês |
| Origem: | Biblioteca Digital do IPB |
| Resumo: | Fall detection systems have traditionally relied on sequential pattern recognition methods, using, for example, time series data obtained from inertial sensors, such as accelerometers. This paper proposes a methodology for fall detection based on converting time series from accelerometer sensors into visual representations using the Markov Transition Field (MTF) method. The UP-Fall dataset was used to test the performance of a Convolutional Neural Network (CNN) model trained on the MTF images generated. A systematic analysis of the image generation parameters was carried out, including the window size, the percentage of overlap, and the number of bins used in the discretizations. The experiments showed that the configuration with 55 bins, a window of 200 samples, and 40% overlap resulted in the best accuracy (97.13%), demonstrating that the conversion of sensory signals into MTF images is a promising alternative for fall detection, allowing computer vision models to capture relevant temporal patterns with high efficiency. |
|---|
Atividades financiadas
Carregando projetos financiados...