Using Gesture Detection as a User Interface for Customized Earphones

Nowadays, Dopple's wireless earphones have touch buttons as a user inter-face, however these take in quite some physical space. Since the earphonesare becoming smaller over the years, there is a need for a new type of touch-less user interface, that is smaller. In order to create that new user interface,research is done on the topic of remote gesture sensing with sensors that cant onto wireless earphones. When gestures are recognized by the system,a corresponding action can be taken, like for example pausing the music.Small infra-red imaging sensors are chosen as a solution to the problem. Itsimages are analysed by a trained image recognition neural network createdwith Python and Keras. This network takes an image as input and outputsa gesture. Each gesture is supposed be linked to an action in the new userinterface. This report focusses on the retrieval of low resolution infra-redimages and neural network training/machine learning. The APDS-9500, analready existing moving gesture sensor, is used as a comparison with regardto the new neural network technique. It achieves an accuracy of 92.3% with5 dierent gestures. The AMG Grid Eye is an 8 by 8 pixel infra-red camerafor which 5 gestures are trained. For raw images, 5 gestures are recognizedwith an accuracy of 79.2%. With the help of pre-processing in the form ofcontrast increasing and linear extrapolation the accuracy is increased up to92.4%. The FLUKE 279 FC is a high resolution camera mounted on a mul-timeter of which the images are downscaled to 30 by 30 pixels. It is foundout that while the size of the model increases, accuracy also increases up to97.2% for 5 gestures. When the FLUKE is tested with 9 dierent gestures,while also optimizing for size, an accuracy is achieved of 98.2% with a modelof 203 kB. The study proves that 30 by 30 infra-red images contain enoughinformation to use gesture recognition with a small neural network.Biomedical Engineerin