Acoustic Sensor Localization Techniques Using Artificial Sound Sources in Reverberant Environments

Abstract

Widespread use of smart devices has brought a growth of user-customized services. In particular, localization techniques have been gaining attention due to increase of location-based services (LBS). Most of LBS services such as navigation systems, traffic alerts or augmented reality (AR) services depend on the GPS for its accuracy and speed, however, its operation is limited to the outdoor environments. The demand of indoor LBS is rapidly growing due to the growth of automated home and IoT technology. There have been studies via WiFi, Bluetooth or RFID, but their performance has been unsatisfactory for their limitation such as the requirement of additional equipment or guarantee of the line of sight.

Among various sensors used for indoor localization, we focus on the acoustic sensors, i.e. microphones. There are several advantages in using the acoustic signals for indoor localization. There is no need for additional apparatus since loudspeakers are pre-installed in most of the buildings for the purpose of announcement or playing background music and mobile devices such as cellphones or tablets are equipped with microphones and loudspeakers. Even the prevailing popularity of IoT services helps accessibility of acoustical sensors and loudspeakers. In addition, acoustic signals have advantages of being able to detect signals through obstacles unlike cameras of RFID.

In this thesis, we propose a position estimation system using acoustic signals to maximize these advantages. We aim to estimate the position of the target user with an acoustic sensor based on the recording of signals from the fixed loudspeakers installed around the room. We target to estimate the position of the acoustic sensor with high accuracy and low-complexity in a large space with high reverberation. Particularly, we try not to affect human hearing by using inaudible frequency bands. In order to estimate the position, it is important to estimate the direct path signal rather than the signal due to reverberation or reflection. To do this, we present various localization techniques as following.

First, we propose the source data structure to operate in the large reverberant environments. In the large space, the consideration of the near-far effect is required which refers to a situation when the desired signal is far away, it is difficult to receive the desired signal due to the interference of closer unwanted signals. In wireless communications, it can be dealt with by interaction of transmitter and receiver by feedback of channel information. However, it is difficult in the acoustic system since there is no feedback between the transmitter and receiver. We borrowed the structure called OFDMA-CDM and modified it to deal with the near-far effect. In the reverberant environment, the amplitude of reverberation is often larger than the direct path signal. We proposed the technique to estimate the direct path signal.

Second, we propose a method for accurate location estimation in the highly reverberant environments. In the high reverberation condition, more spurious reflections occur, which makes it difficult to estimate the time delay of the direct path signal. If the time delay estimation is wrong, it is likely that the position estimate does not converge by an estimation method. In the proposed method, position candidates are obtained from most of the received signals including signals even from spurious reflections. The unreliable candidates are filtered out by the agreement test and rank the rest candidates by their reliability to find accurate target position. We can estimate the receiver's position even in the condition of attenuated direct path signal or high reverberation by using the proposed method.

Third, we proposed a low-complexity localization method to work in the highly reverberant environment. This method is based on the particle filter that estimates the position by weighted particles whose weights are computed by the likelihood. We designed likelihood function that efficiently calculates likelihood in the region with the direct path signal so that more reliable position can be obtained. The proposed method enables location estimation with high precision with a relatively small amount of computation in severe reverberation.

The proposed methods are evaluated in simulated environments with different reverberation time. The performances are verified in different parameters and compared with other localization methods. In addition, the performance is evaluated in the real reverberant environment with a large space. A series of experiments has shown the superiority of the proposed methods and it is appropriate to apply in the actual environment.