Acoustic event filterbank for enabling robust event recognition by cleaning robot

Abstract

Due to its mobile capability when performing house-cleaning function in absence of home owners, a cleaning robot has sufficient capacity to be fully utilized as an automatic surveillance system for indoor security. While many research efforts have been made recently to provide a robot understand the auditory environment, there are still many obstacles to overcome. One of the most serious challenges encountered in providing accurate auditory scene analysis is the presence of robot ego noise. Robot ego noise is primarily generated by the embedded motors on a robot during its operation. This paper proposes a new filterbank design based on discriminative distances within event-to-noise and event-to-event, respectively. The proposed filterbank essentially is designed to provide reliable recognition under cleaning robot ego noise through the result of event spectrum analysis. The experimental results indicate that the features extracted by the proposed filterbank are more robust than the conventional ones.