|
Real-Time Fall Detection for Elderly Care Using YOLOv8 with a Custom-Built Image Dataset |
|---|---|
| รหัสดีโอไอ | |
| Creator | Supakorn Ukampan |
| Title | Real-Time Fall Detection for Elderly Care Using YOLOv8 with a Custom-Built Image Dataset |
| Contributor | Peerapong Phanthura, Niwat Angkawisittpan, Bin Zheng, Somchat Sonasang, Thipwimon Chompookham, Worawat Sa-Ngiamvibool, Taweesak Thongsan |
| Publisher | Faculty of Engineering Mahasasakham University |
| Publication Year | 2568 |
| Journal Title | Engineering Access |
| Journal Vol. | 11 |
| Journal No. | 2 |
| Page no. | 271-276 |
| Keyword | fall detection, YOLOv8, elderly care, computer vision, real-time monitoring |
| URL Website | https://ph02.tci-thaijo.org/index.php/mijet/index |
| Website title | THAIJO Engineering Access |
| ISSN | 2730-4175 |
| Abstract | This paper aims to develop a model for human fall detection by simulating authentic fall incidents for implementation in a computer vision system designed to monitor falls in the elderly and deliver real-time notifications. The model development process commences with the utilization of a dataset comprising item bounding boxes and corresponding annotations. The YOLOv8 methodology is subsequently employed to train the dataset. The study dataset consists of 2,788 raw images that have been annotated and processed using Roboflow technology. The images are categorized into three groups: the training set comprises 77% of the data, totaling approximately 2,146 images; the validation set constitutes 12%, or about 338 images; and the test set accounts for 11%, roughly 304 images. Data augmentation methods were used in the fourth stage of the Roboflow platform to increase data diversity, resulting in 19,000 images. This expanded dataset enhances the model's ability to generalize by exposing it to a wider variety of scenarios and conditions. Consequently, the increased volume of images allows for more robust training, ultimately improving the accuracy and reliability of the model's predictions in real-world applications. The ideal value for improving model performance is one hundred epochs, which is how long model training was run. The model testing outcomes, carried out in the same setting as the training, show a mean average accuracy (mAP) of 90.97% and an overall accuracy of 95.36%, suggesting outstanding accuracy and appropriateness for practical use. |
