Loading...
Please wait, while we are loading the content...
Similar Documents
Optical myography system for hand posture and gesture recognition
| Content Provider | Semantic Scholar |
|---|---|
| Author | Serpa, Alberto Luiz Hélio |
| Copyright Year | 2018 |
| Abstract | In this work, an optical myography system is demonstrated as a promising alternative to monitor hand posture and gestures of the user. This technique is based on accompanying muscular activities responsible for hand motion with an external camera, and relating the visual deformation observed on the forearm to the muscular contractions/relaxations for a given posture. Three sensor designs were proposed, studied and evaluated. The first one intended to monitor muscular activity by analyzing the spatial frequency variation of a uniformly distributed stripe pattern stamped on the skin, whereas the second one is characterized by reckoning visible skin pixels inside the region of interest. Both designs are impracticable due to their low robustness and high demand for controlled experimental conditions. At last, the third design retrieves hand configuration by tracking visually the displacements of a series of color markers distributed over the forearm. With a webcam of 24 fps and 640 × 480 pixels, this design was validated for eight different postures, exploring fingers and thumb flexion/extension, plus thumb adduction/abduction. The experimental data are acquired offline and, then, submitted to an image processing routine to extract color and spatial information of the markers in each frame; the extracted data is subsequently used to track the same markers along all frames. To reduce the influence of human body natural and inherent vibrations, a local reference frame is yet adopted in the region of interest. Finally, the frame by frame data, along with the ground truth posture, are fed into a sequential artificial neural network, responsible for sensor supervised calibration and subsequent posture classification. The system performance was evaluated in terms of eight postures classification via 10-fold cross-validation, with the camera monitoring either the underside or the back of the forearm. The sensor presented a ∼92.4% precision and ∼97.9% accuracy for the former, and a ∼75.1% precision and ∼92.5% accuracy for the latter, being thus comparable to other myographic techniques; it also demonstrated that the project is feasible and offers prospects for human-robot interaction applications. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://repositorio.unicamp.br/bitstream/REPOSIP/331230/1/Wu_YuTzu_M.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
