Personalized Similarity Models for Evaluating Rehabilitation Exercises from Monocular Videos
| Authors | |
|---|---|
| Year of publication | 2024 |
| Type | Article in Proceedings |
| Conference | 17th International Conference on Similarity Search and Applications (SISAP) |
| MU Faculty or unit | |
| Citation | |
| Web | https://link.springer.com/chapter/10.1007/978-3-031-75823-2_7 |
| Doi | http://dx.doi.org/10.1007/978-3-031-75823-2_7 |
| Keywords | pose estimation;skeleton sequence;rehabilitation exercise;human body keypoint;exercise quality assessment;exercise similarity;personalized similarity;kNN retrieval |
| Attached files | |
| Description | Automatic monitoring of exercise correctness during home physical rehabilitation could significantly increase the impact of rehabilitation treatments. To evaluate exercise quality effectively, it is necessary to extract relevant spatio-temporal motion features and compare them to an ideal exercise pattern. We argue that the features should be personalized to the patient's needs, as the movement abilities of each patient are specifically limited and also change over time. Towards this end, we utilize the MediaPipe Pose tool to estimate 2D and 3D coordinates of skeleton joints from a monocular video stream. The joint coordinates are then processed to extract specific spatio-temporal features that are automatically weighted for each patient. This allows for personalized similarity based on the individual's exercise patterns while requiring minimal training data and possibly offering explainable evaluations. The proposed approach is tested on the REHAB24-6 rehabilitation dataset, reaching superior effectiveness and being about 2-3 orders of magnitude more efficient than state-of-the-art solutions. |
| Related projects: |