Title

Acoustic Echo Estimation using Model-Based Approach with Application to Spatial Map Construction in Robotics

Assessment committee

Dr. Akihiko K. Sugiyama, Yahoo! Japan

Professor Kazuhiro Nakadai, Tokyo Institute of Technology

Associate Professor Claus B. Madsen (committee chair), Aalborg University

Supervisors

Associate Professor Jesper Rindom Jensen, Aalborg University

Abstract

Constructing an accurate map of an indoor environment is an active area of research withinrobotics. Camera- and laser-based technologies are commonly used to generate a spatial map ofan environment. These maps are used to enable robots to move within an environment. However, these modalities have limitations as these technologies cannot detect transparent surfacestypically found in an office environment. Moreover, camera- and laser-based technologies sufferfrom a limited field of view, limiting the ability to generate a spatial map. These limitations canbe addressed by utilizing sound. This inspiration comes from animals that utilizeecholocationto make “sense” of their environment, i.e., get spatial information about the environment using sound. Animals such as bats, rats, and dolphins are among the few that have mastered the art ofecholocation with accurate precision. The main research question that we attempt to answer inthis thesis is that if animals in nature can use echolocation for navigation, can we enable robotsto utilize echolocation to navigate an environment?

The central theme of this thesis is to propose the utilization of the concept of echolocationand combine it with advanced audio processing techniques that can complement existing robotperception technologies to estimate acoustic echoes. To this end, we propose a model-basedapproach to detecting and estimating acoustic echoes, i.e., we utilize a sound propagation modelto formulate the problem of estimating a parameter of interest (POI), e.g., time of arrival (TOA)of acoustic echoes. Therefore, two methods to resolve the acoustic echo estimation problemare presented in this work: a non-linear least squares (NLS) estimator and an Expectation-Maximisation (EM) approach. Both methods estimate TOA/DOA directly from the observed
signals. In this thesis, we first propose a single-channel TOA estimation technique using theNLS and EM methods. Later, these methods are extended into the multi-channel approach toestimate the direction of arrival (DOA) information of an acoustic reflector by estimating thetime difference of arrival (TDOA) of acoustic echoes. Here, spatial filtering techniques, e.g., beamforming techniques, are utilized to localize the position of an acoustic reflector. Delayand sum beamformer (DSB), minimum variance distortionless response (MVDR) beamformer, and linear constraint minimum variance (LCMV) beamformer are used to estimate the DOAs of
acoustic echoes. A proof of concept (POC) robotic platform was built to test the performance of the proposed methods. The estimators and beamforming techniques were implemented to acquire actual data and later used to localize acoustic landmarks for spatial map construction.

Finally, the detection problem was extended, and a new approach was proposed, where weutilize the ego-noise of a robotic platform to detect the presence of an acoustic landmark. Thiswork paves the way for a novel sound-based collision avoidance system that gives 360o of spatial awareness which could be utilized in robots and drones.This thesis will begin by reviewing the overall architecture of the robotic platform and later narrow down the discussion to robotic perception. The problems associated with estimatingacoustic reflectors, e.g., walls, using a traditional approach, e.g., cameras and lidar, are alsodiscussed. Discussion related to utilizing echolocation for acoustic spatial map constructionis also an important highlight of this thesis. Based on the discussion in Section 1, a researchquestion is formulated and later the aims and objectives of this thesis are discussed. In Section2, the acoustic echo model is presented based on which we derive the nonlinear least squares (NLS) and expectation-maximization (EM). Discussion related to estimating distance, directions of acoustic echoes, and representing this information to construct an acoustic map of an environment is found in Section 2 and Section 3. The contribution is found in Section 4 while the conclusion and future work are discussed in Section 5.