Computer vision based patient activity monitoring systems can be attractive for various unobtrusive clinical applications. Such a monitoring system can be developed using movement information derived from the skeleton model of the current body pose, e.g. obtained using a depth camera. Earlier research using estimated skeleton models have been focused mostly on gaming applications. In this paper, we propose CNN-SkelPose as a skeleton model estimation method for clinical applications. CNN-SkelPose uses a trained Convolutional Neural Network to extract both the local and global information from the depth image. CNN-SkelPose outperforms the baseline model of Skeltrack for reliable skeleton model estimation in patient monitoring scenarios. Our results show the inadequacy of existing methods for skeleton model estimation when applied to a clinical scenario and suggests CNN-SkelPose as an improvement towards this application.
|Number of pages||12|
|Journal||Journal of Ambient Intelligence and Humanized Computing|
|Early online date||28 Feb 2019|
|Publication status||Published - 1 Jun 2020|
- Clinical application
- Skeleton model