Riconoscimento di gesti tramite sensore wearable per applicazioni in robotica

Broggini, Denis (2017) Riconoscimento di gesti tramite sensore wearable per applicazioni in robotica. Bachelor thesis, Scuola Universitaria professionale della Svizzera italiana (SUPSI).

[img] Text

Download (316kB)


The aim of this project has been to insert an artificial intelligence component in the current IDSIA’s human-robot interfacing system, allowing an operator with wearable sensors to change the behavior of a neraby robot via a robust detection of target gestures. First of all we defined and analyzed the two gestures to recognize: the ’stop gesture’, which is the first to be detected in order to familiarize with the technologies, and the ’pointing gesture’, of which the robust recognition is the final goal of the project. Then, in a second phase we recorded a series of sessions, where the operator, wearing the sensors and acting like in a realistic work environment, at the occurring of some specific sounds performed the gestures of interest. Using this event-based approach and the sliding window algorithm on the logged data, it has been possible to determine in a post processing phase which data was representing these gesture and which not. With this dataset we trained a series of classifiers and compared according to IDSIA’s specific purpose. The classifier with the best results has been a Neural Network using input windows of sensor data belonging to the last 2.5 seconds, with a true positive gesture recognition rate of 100%, and less than one false positive recognition per minute. In general, the obtained results are encouraging. We made a generic process of acquisition and learning, so it is ready for new gestures. The final product works in real-time and has reasonably good results with different persons.

Item Type: Thesis (Bachelor)
Supervisors: Gambardella, Luca Maria and Giusti, Alessandro and Gromov, Boris
Subjects: Informatica
Divisions: Dipartimento tecnologie innovative > Ingegneria informatica
URI: http://tesi.supsi.ch/id/eprint/1772

Actions (login required)

View Item View Item