1.1 Motivation and Objective
Robots are getting closer to human life in recent years. The interaction between robots and human is a key factor how the robot could be applied to real environment with human around. Among the interaction scenarios between robot and human, localization of the robot is frequently discussed. Lots of the localization methods that have been studied are using camera or laser range finder as the sensor input.
The cameras are able to distinguish different landmarks by using the visual information such as color, object edge, shape of landmarks, etc. The association between different temporal frames of sensor data is feasible. On the other hand, laser range finders are able to provide accurate range information of different landmarks. A more precisely measured data would help the localization procedure converges faster.
However, both of these two sensors are suffering from various drawbacks such as occlusion. If a landmark is NLOS (non-line-of-sight), it will be not detectable.
We intend to propose a method focusing on solving the occlusion problem.
Acoustic signals are detectable even if a landmark is NLOS. Also, in human environment, there are stationary sound sources such as air-conditioner, operating sound of a computer, and non-stationary sound as human speech, door-knock, etc.
According to this characteristic, we consider multiple sound sources as the landmarks in the localization problem. In that sense, a high-dimensional microphone array is mounted on a mobile robot to detect these sound signals, and the robot has its own encoder information to help the self localization of the robot.
1.2 Literature Review
The array signal processing technology [1] was first used in the World War I. The invention was used to detect enemy aircrafts. The technology was applied to array telescope such as the Very Large Array in New Mexico, USA. In this thesis, we use a microphone array to implement the array signal processing algorithm [2].
There are generally two categories of method to solve direction of arrival (DOA) estimation problem. One of the methods is TDE (time delay estimation) proposed by Knapp and Carter [3]. Two microphones are used to record the sound signal, and sound sources from different direction will cause different response to these two microphones. To measure the time delay between microphones, a commonly used method is GCC (Generalized Cross Correlation). Finding the maximum value of the GCC expression indirectly indicates knowing the delay relation. The temporal difference between these two responses could be used to estimate the direction.
The second well-known DOA estimation method is eigenspace method. It measures the distribution of eigenvector between different signals and estimates the signal direction by mutual projection. The MUSIC algorithm proposed by Schmidt [4]
belongs to this category. It is able to localize multiple sound sources with prior knowledge of the total number of sources.
After the DOA estimation, we use this information to localize the mobile robot.
There are a lot of researches focusing on solving SLAM (simultaneous localization and mapping) problem. Bayes filters [6] are commonly used to both describe the position state of a robot itself and the environmental landmarks position estimation.
1.3 Thesis Subject and Contribution
The subject of this thesis can be divided into two parts. The first part is to implement a multiple sources DOA estimation algorithm. The DOA estimation results will be considered as the bearing measurements of the second part of the thesis. The second part of the method is the Bearings-Only SLAM algorithm that is able to simultaneously localize the features in the experimental environment and to map the robot itself to the environment.
In the first part, we compare the effectiveness of the two algorithms, which are ES-GCC and accumulative MUSIC. ES-GCC is an eigenspace based method that is able to detect unknown number of multiple sound sources direction simultaneously.
MUSIC is a more commonly used DOA estimation method that could only estimate known number of multiple sound sources. We monitor the number of time frames needed for both algorithms in real time application, and modified them to satisfy the hardware limitation.
The second part is a SLAM problem architecture, which considers the outputs of the first part as the measurements. SLAM problem could be solved using procedures based on the Bayes filter. The particle filter is used for the non-parametric robot environment. The removal of the resample step in particle filter doesn’t disturb the algorithm but decrease the computing complexity and shorten the algorithm procedure so that the method is able to be applied in real time.
The experimental results are shown to justified the valid modifications, yet accomplishes the simultaneous localization and mapping problem in real-time.
1.4 Outlines of Thesis
The remainder of this thesis is organized as follows.
Chapter 2: The two kinds of DOA estimation method are clearly described, including the performance analysis for real time application. The reasons of choosing eigenspace method are explained and there would be modification made to the algorithm to solve the insufficient performance of real time experiment.
Chapter 3: The detail concept of Bayes filter is stated in this chapter. Based on the concept, a bearings-only SLAM algorithm constructed by EKF (extended Kalman filter) and PF (particle filter) is introduced. The mathematical detail is also in this chapter. Finally, PF is able to deal with the unknown data association between different temporal frames.
Chapter 4: The experimental results are presented. A combinational architecture of DOA estimation and bearings-only SLAM is applied to a real mobile robot, and the real time performance analysis is discussed.
Chapter 5: The conclusion of this thesis and the possible improvement in the future is presented is this chapter.