Vision-based fall detection system using deep learning on an edge-cloud architecture | |
| Author | Tonson Praphabkul |
| Call Number | AIT Thesis no.DSAI-24-05 |
| Subject(s) | Human activity recognition Deep learning (Machine learning) |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Data Science and Artificial Intelligence |
| Publisher | Asian Institute of Technology |
| Abstract | Fall detection is a subactivity from the Human Activity Recognition (HAR) task. This approach involves having a machine recognize an action that a person has completed and categorize it into fall and non-fall categories. There are many ways to detect falls. One way is to use the signal input from the measuring device, such as an accelerome ter or gyroscope, and pass it to 1D-CNN with the LSTM model. The other way is to use an image sensor and pass a stack of RGB images into a CNN-based model. Cur rently, the state-of-the art model is a 3D-CNN-based architecture. This model uses a sequence of skeleton heatmaps. One of the drawbacks of a 3D-CNN-based architecture is its inference speed due to its computational complexity. Based on the state-of-the-art model, I designed a new model that is more effective while maintaining the ability to classify the action. This was achieved by finding the optimal amount of residual block through experiments. On the NTURGB+D dataset, the downsize model performance is almost equivalent to the current state of the art model, attain top1 accuracy of 92.32% and top5 accuracy of 99.51% for activity profiling tasks. For the fall detection task, the downsize model have almost equals the performance of the most advanced model, with 100% accuracy, precision, recall, and F1-score. For the SAFER dataset, the results are more or less similar. For the SAFER non-wheelchair set, the model achieves higher ac curacy but lower precision and recall. In the SAFER wheelchair set, the accuracy is the same: higher in precision but lower in recall. For fall detection in the SAFER dataset, the downsize model gets around 83 % accuracy, while precision and recall are around 90% for both sets. |
| Year | 2024 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Data Science and Artificial Intelligence (DSAI) |
| Chairperson(s) | Mongkol Ekpanyapong |
| Examination Committee(s) | Huynh, Trung Luong;Chaklam Silpasuwanchai |
| Scholarship Donor(s) | Royal Thai Government |
| Degree | Thesis (M. Sc.) - Asian Institute of Technology, 2024 |