Hand Position and Movement Acquisition System for a Wrist Based HMI
Fasel, Daniel (2021)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2021052711612
https://urn.fi/URN:NBN:fi:amk-2021052711612
Tiivistelmä
The field of human-machine interfaces keeps evolving at a steady pace. In recent years gesture recognition has become an important focus of research. Most commercial applications of gesture recognition work with the help of cameras. These suffer from occlusion and low portability. Wrist based systems do not suffer from these problems, making them ideal for use cases such as translating sign language.
Current research on wrist based devices rely on neural networks to recognise the gestures. They mostly use an arbitrary set of gesture to label their data. This limits the applications of such devices as the fine movements of the hand cannot be classified by the neural networks. Any additional gesture also requires new training that is bound to the sensors used. A real-time hand position and movement acquisition system was developed in this work to improve the labeling of sensor data. With such a system researchers can create a neural network that uses the exact position of the hand instead of an arbitrary set of gestures. This device takes the shape of a 3D printed exoskeleton placed on the back of the hand. Potentiometers are used to measure the angle of each joint of the hand.
The exoskeleton has good anatomical fidelity, is affordable and easy to build. The one sigma standard deviation of the measurement accuracy is 3.1 degrees. This relatively poor accuracy is due to the potentiomteters. Individual calibration of the sensors could improve this result. This work demonstrates that using 3D printing and potentiometers is a viable solution to capture the position of the hand in real-time.
Current research on wrist based devices rely on neural networks to recognise the gestures. They mostly use an arbitrary set of gesture to label their data. This limits the applications of such devices as the fine movements of the hand cannot be classified by the neural networks. Any additional gesture also requires new training that is bound to the sensors used. A real-time hand position and movement acquisition system was developed in this work to improve the labeling of sensor data. With such a system researchers can create a neural network that uses the exact position of the hand instead of an arbitrary set of gestures. This device takes the shape of a 3D printed exoskeleton placed on the back of the hand. Potentiometers are used to measure the angle of each joint of the hand.
The exoskeleton has good anatomical fidelity, is affordable and easy to build. The one sigma standard deviation of the measurement accuracy is 3.1 degrees. This relatively poor accuracy is due to the potentiomteters. Individual calibration of the sensors could improve this result. This work demonstrates that using 3D printing and potentiometers is a viable solution to capture the position of the hand in real-time.