机器人声源定位技术综述：麦克风阵列方法

需积分: 9 82 浏览量更新于2024-07-17 1 收藏 1.36MB PDF 举报

"这篇文档是关于机器人领域中声源定位技术的综合回顾，重点讨论了麦克风阵列在声源测向中的应用。" 在现代机器人技术中，声源定位（Sound Source Localization, SSL）是一个至关重要的研究领域，特别是在自主导航、环境感知以及交互性任务中。这篇文献"17J_Localization of sound sources in robotics A review_麦克风阵列声源测向入门综述.pdf"提供了一个详尽的概述，深入探讨了SSL在机器人平台上的应用及其挑战。文章首先对SSL技术进行了分类，包括基于到达时间差（Time Difference of Arrival, TDOA）、到达角度（Direction of Arrival, DOA）和能量强度等多种方法。这些技术允许机器人通过多麦克风阵列来识别声音的来源，从而实现对周围环境的听觉理解。 SSL问题的描述涉及多个方面：除了确定声源的方向之外，还包括距离估计和跟踪。TDOA方法通过比较不同麦克风接收到信号的时间差异来确定声源位置，而DOA方法则分析信号到达不同麦克风时的角度变化。距离估计是通过声波传播速度和到达时间来计算的，这在机器人定位和避障中尤其有用。至于跟踪，机器人需要连续更新声源的位置信息，以适应移动声源或机器人自身运动的情况。论文还详细介绍了评估SSL性能的各种方法，包括仿真测试、实验室实验以及实地试验。这些评估方法对于衡量不同算法在实际环境中的表现至关重要，帮助研究人员优化和改进SSL技术。当前的SSL挑战主要包括噪声抑制、多声源区分、实时处理能力和计算效率等。在嘈杂环境中准确地分离并定位声源是一项极具挑战性的任务，尤其是在存在多个同时发声的声源时。此外，SSL系统需要具备实时处理大量数据的能力，并且要在有限的计算资源下运行，这对算法的设计提出了高要求。最后，文献中提出了未来的研究方向，如深度学习在SSL中的应用、增强现实与SSL的结合以及多模态感知对提高SSL精度的影响。这些前沿话题展示了SSL技术在未来机器人系统中的巨大潜力和发展空间。这篇综述提供了丰富的SSL技术背景和实践案例，对于想要深入了解麦克风阵列在机器人声源定位中的应用的研究者和工程师来说，是一份非常有价值的参考资料。

C. Rascon, I. Meza / Robotics and Autonomous Systems 96 (2017) 184–210 189

the final θ

is the one associated with the IPD

from the set that is

the most similar to the estimated



IPD

Alternatively, machine learning approaches can be used to

tackle SSL. These approaches do not define a sound propagation

model, instead they learn a mapping function Ψ from a space of

features κ of the sound signal to source locations θ.

: κ ↦→ θ . (11)

This map is modeled from recorded examples (training dataset)

that are used to identify the parameters φ. Traditionally, additional

examples are recorded to evaluate the trained model (testing

dataset), and is expected to perform well in a real-world setting.

Some types of training techniques that have been used for SSL are:

recurrent neuronal networks [80–83]; bio-inspired spiking neural

networks [84]; and deep learning architectures are recently being

explored [85–87].

However, in order to train a model, training data diversity

is important for generalization [88]. This is because the training

process requires to ‘‘observe’’ a representative set of examples

from which it will attempt to generalize the whole set of possible

scenarios. If the circumstances in which the model is deployed in

a real-world setting are vastly different from the circumstances in

which it was trained, a basic assumption of most machine learning

techniques is overlooked [89]. For instance, if the training data

was recorded in a quiet office and the robot is deployed in a busy

restaurant, the mismatch in noise levels will result in a poor SSL

performance. To avoid this, a highly diverse training dataset is

required, varying in terms such as: room response, number and

location of sources, microphone placement, the use of external

pinnae and/or inner-ear canals, head model, etc. In [90], a system

is presented that both locates and identifies the users, and the

authors proposed as future work to evaluate their SSL performance

in mismatched training and testing conditions. An example of how

this phenomenon impacts the performance of a SSL system can be

found in [57], detailed in Section 4.2.

4.2. 2-dimensional single direction-of-arrival estimation

In a 3D environment, it is of interest to estimate the DOA of the

source in both the azimuth and elevation plane. If the microphone

array has a 3D geometry, the elevation angle can be calculated us-

ing the same TDOA-based techniques discussed in Section 4.1 along

with the DOA in the azimuth plane, providing a 2-dimensional DOA

estimation. This requires, however, additional microphones to be

used.

Frequently in binaural hearing the microphones are positioned

with a body between them (breaking the free-field assumption)

and/or accompanied with external ears or with inner-ear canals.

This is carried out so that the sound source signal is ‘filtered’ before

it is captured by the microphones in a way that is dependent of its

direction in both planes. Even though the free-field assumption is

broken, this filtering effects can be used alongside non-free-field

propagation models to carry out 2-dimensional DOA estimation.

For example, human beings are able to estimate both the az-

imuth and elevation of a sound source even when employing a

two-microphone array [91]. When a sound source located on the

side of a human head emits a sound wave, the ear farthest away

from the source receives a modified version of the one received

by the closest ear. This modification is carried out by several

measurable physical phenomena when passing through and/or are

reflected by

the human head, torso, external pinnae and inner-

ear canals [92]. This set of phenomena can be measured by placing

The physical phenomenon observed when the audio signal is passing through

the body are generally present in low frequencies, while the ones reflected are

present in high frequencies.

a microphone on each side of a specialized dummy head and

capturing their impulse responses in an anechoic room (to avoid

capturing the effects of the environment). These measurements

can be used to calculate how the object between the microphones

changes the properties of the signal received, depending on the lo-

cation of the sound source (specifically, its azimuth and elevation).

This set of properties can be used to create what is known as a

head-related transfer function (HRTF), which is a type of filter that

aims to emulate the physical phenomena that modify the sound

source audio wave, given a pre-specified 2-dimensional DOA [93].

An HRTF can be used, among other applications, for spatial au-

dio reproduction where a sound source is virtually ‘‘positioned’’

around a listener wearing headphones [94]. For the purpose of

single-DOA estimation, a set of candidate 2D DOAs is proposed,

with which a set of HRTFs are measured. Then, a database of filters

that carry out the inverse function of an HRTF (referenced here

as IHRTF) can be calculated from each HRTF-to-DOA association.

These IHRTFs can then be applied to the incoming signals following

a grid-search mapping. From each application, a proposed metric

can be measured, such as the correlation between the output

signals. The DOA associated with the IHRTF that maximizes such

metric is then proposed as the estimated DOA [95]. It is important

to consider that the external pinnae can also play an important role

in the estimation of the HRTF, as shown in [96] where a spiral ear

is used for 2-dimensional DOA estimation.

However, this type of approach requires measuring the ef-

fects of the database-IHRTFs (which are based on HRTFs usually

estimated with measurements in low-noise anechoic rooms) to

captured signals in real-world conditions (which usually include

reverberation and noise). To counter this, room characteristics

could be measured in advance and be considered as part of the

calculations of the database-filters (philosophically different from

the aforementioned IHRTFs), making them consistent with the

real-world conditions. Unfortunately, deploying a robot into a real-

world environment implies that the acoustic characteristics are not

known in advance [97]. In addition, the room response is depen-

dent of the position of the microphone array inside the room. This

means that when a robot moves (either linearly or by rotation), a

new room response must be measured and a new filter database

must be calculated to account for this change [98]. A simulation

of such a real-world environment can be carried out [99,100],

from which room characteristics can be obtained to automatize

the filter-database modification. However, carrying out such a

simulation can be unfeasible since the dimensions and materials of

the environment may not be known in advance. In the cases where

they are known, such a simulation is time-consuming [44].

The revised auditory epipolar geometry (RAEG) [44], described

in Section 4.1, can be used to surmount the issue of differences

between real-world and training conditions. RAEG-based SSL does

not require a set of IHRTFs to be estimated in an anechoic room,

since the shape of the robotic head is already accounted for. How-

ever, it only provides a DOA in the azimuth plane. A generalization

to 2D localization is possible via scattering theory.

Scattering theory is a field of physics which models how a wave

or particle is perturbed from its path by some object [101,102].

In the case of acoustic or electromagnetic waves, this modeling

process can be used for source localization [103]. As shown in [97],

scattering theory, with a given IPD and IID, can be used to calculate

both azimuth and elevation DOAs by assuming that the shape of

the object between the microphones is spherical. A grid-search

mapping (similar to the IHRTF-based and RAEG-based approaches

already described) is used in [97] with a close-to-spherical robotic

head between the microphones. The approach is summarized as

follows: (1) a set of [IPD

, IID

] tuples is calculated for each possible

combination of azimuth DOA, elevation DOA and f ’s using the

scattering theory equations; (2) an



IPD

and



IID

tuple is estimated

剩余26页未读，继续阅读

潮浪之巅

粉丝: 4w+
资源: 1

机器人声源定位技术综述：麦克风阵列方法

localization-of-sound-source.zip_localization_sound localization

robot_localization_localization_雷达定位ekf_环境建图_ROS_ros小车建图_

ＲＯＳ定位功能包：ｒｏｂｏｔ＿ｌｏｃａｌｉｚａｔｉｏｎ

error: cannot launch node of type [robot_localization/ekf_localization_node]: robot_localization

ros仿真报错ERROR: cannot launch node of type [fake_localization/fake_localization]: fake_localization

如何使用robot_localization功能包

robot_localization使用教程

ubuntu24安装robot_localization

怎么录制basic_localization_stage_indexed.bag

安装robot_localization

最新资源