Improved Sensor Planning in Binaural Probabilistic Active SoundLocalisation for Robots
Abstract: Sensor planning in the field of active robot sound localisation is an important topic that has potential applications for robots interacting with crowds, search and rescue and more, though the field is currently at the level of basic research. This thesis presents novel improvements that theoretically increase the accuracy of the predictions used in probabilistic sensor planning for sound localisation. In particular two novel improvements are made: First, N-step ahead sensor planning using dierential entropy with a full Gaussian Mixture Model (GMM) is developed and compared with only approximating the prediction with a single Gaussian. The results indicate a modest increase in short-term performance when using a full GMM, but worse or comparable performance in the long term. Secondly, the eect of dierent observation models for the predictions are evaluated to determine the optimal choice. The result of this is surprising, showing that the simplest model performs the best.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)