^{1}

^{2}

^{3}

^{1}

^{2}

^{3}

^{1}

^{2}

^{*}

^{1}

^{2}

^{3}

^{1}

^{2}

^{1}

^{2}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (

In this paper, a MEMS microphone array system scheme is proposed which implements real-time direction of arrival (DOA) estimation for moving vehicles. Wind noise is the primary source of unwanted noise on microphones outdoors. A multiple signal classification (MUSIC) algorithm is used in this paper for direction finding associated with spatial coherence to discriminate between the wind noise and the acoustic signals of a vehicle. The method is implemented in a SHARC DSP processor and the real-time estimated DOA is uploaded through Bluetooth or a UART module. Experimental results in different places show the validity of the system and the deviation is no bigger than 6° in the presence of wind noise.

Direction finding of moving vehicles by microphone arrays is very important in unattended ground sensor (UGS) systems [

The bearing of vehicle is an essential piece of intelligence and could also provide assisting information for other sensors. Direction finding is the basis of vehicle detection [

To design a real-time direction finding system, it is very important to choose a suitable DOA estimation method. The criteria for choosing the method are given below:

Low complexity for real-time processing

High accuracy for the performance of the system

Moderate sampling rate for the hardware load

In general, methods for acoustic source direction finding can be divided into three categories based on their increasing computational complexity: time-delay-based methods [

Another challenge for a microphone array in the field is the wind noise. In this paper we propose a spatial coherence-based method to estimate the useful band for vehicle direction finding. The sound of the vehicle in the field has free field characteristics and the wind noise has the characteristics of a noise field. According to reference [

In this paper, we design and implement a vehicle direction finding system using four MEMS microphones, a SHARC DSP processor, MAXIM simultaneous-sampling ADCs and supplemental hardware circuits. The real-time estimated DOA could be reported through a Bluetooth or UART module. The interference of wind noise in the field is reduced through estimation of the useful frequency band by spatial coherence. Because the designed aperture of the array is small and the acoustic signal of the vehicle is band limited, we use the MUSIC algorithm for its relatively low complexity and high accuracy.

The remainder of this paper is organized as follows: Section 2 presents the hardware design of the microphone array. Section 3 elaborates the signal processing method and software design. It illustrates the direction finding method and the solution to wind noise. System verification and experimental results with MEMS microphone array are given in Section 4 and conclusions are presented in Section 5.

In this section, we first elaborate our choice of the microphone array geometry, and then describe the design of system architecture.

The number of microphones in the array and the array aperture are determined by the following requirements:

The array must have the same resolution in all directions

The vehicle signal occupies the frequency band from 100 Hz to 3,000 Hz [

The microphone array system should achieve high accuracy

In general, uniform circular arrays have the same resolution in all directions and the uniform array could provide enough space for circuit design. Furthermore, to satisfy the spatial sampling criterion

To simplify the complexity of the system design, we decided to use no more than four microphones. The expected accuracy of the direction finding system is less than 6°. To determine the number of microphones and the aperture of the array, different microphone arrays were designed (

The 10 dB and 20 dB level experiments are conducted by 500 Monte Carlo simulations. The room experiments are conducted using the microphone arrays shown in

The block diagram of the prototype MEMS microphone array system is depicted in

As shown in

We first establish here the notation used before describing the direction finding method.

The text in

Let

Let

Let

Let

The sampling rate of the system is 8,192 Hz. To ensure the accuracy of spatial coherence and direction finding, 1,024 samples (1/8 s) are used for calculating the spatial coherence and DOA estimation. One second is divided into two parts. As shown in

Wind noise is the most common interference outdoors. The wind turbulence on the microphone is comparatively incoherent and its speed is much slower than that of sound [

The wind noise occupies a relatively lower frequency band compared to the vehicle sound

Coherence can serve as a criterion to separate the wind noise and the vehicle bands

Spatial coherence is a similarity indicator for signals in the frequency domain. It describes the coherence between two measures at two locations [

Taking FFT time duration

In our case, T = 1/8 s (1024 samples), D = 8.31 × 10^{−5} s (array aperture of 4 cm):

We use 1/8 s in one second to estimate the spatial coherence of the frequency band. _{xy}

The car passes the microphone array between 16 s and 22 s. Spatial coherence is depicted in

The MUSIC estimator is used to compute a directional spectrum in this paper. In some application, the acoustic signal of vehicle is considered as wideband. However, when the microphone array is small, the sound of a vehicle could be viewed as a narrowband signal [

The MUSIC algorithm is based on the fact that the array manifold _{0}) and the noise eigenvectors _{N}

The array manifold changes as the frequency varies, while the decrease of the array aperture will make the change of array manifold smaller. In other words, the error caused by frequency dispersion declines as the array aperture becomes smaller. In this paper, as the aperture of the array is as small as 4 cm and the acoustic signal of vehicle is limited, the error of DOA estimation caused by frequency change in array manifold is negligible. With spatial coherence limiting the signal band, we use _{0} = 2_{L}_{H}

In slot 1 of

STEP-0 Calculate the spatial coherence of signals from the first two microphones, and choose the useful frequency band of [_{L}_{H}

In slot 2–8 of

STEP-1 Collect L (1024) samples of data from the small aperture array of M sensors (4 microphones).

STEP-2 Calculate the Fourier transforms

STEP-3 Construct the covariance matrices _{L}_{H}

STEP-4 _{N}_{0}) (_{0} = 2_{L}_{H}_{MUSIC}_{0}_{N}^{2}]^{−1}

Our method differs from the narrowband MUSIC algorithm [_{L}_{H}_{0} = 2_{L}_{H}

Experimental studies were performed from June 2012 to December 2013 on Chongming Island, Zhoushan Island (the third and fourth biggest islands in China) and a suburban district around Shanghai to demonstrate the feasibility of the system and the direction finding method proposed in this paper in the field. In

Different kinds of vehicles are used as targets for direction finding. Sorting the vehicles by ascending sound pressure level (SPL), the order is as follows: electric bicycle, car, bus, truck, tracked vehicle. The UGS works under different weather conditions within its area of operation, therefore the wind scale and range of direction finding is provided. The SPL, wind scale and range of direction finding reflect the signal to noise ratio (SNR). As for each target, the maximum wind level in the test and range of direction finding are different. For the relative conditions of different vehicles, the noise is a minimum 5 dB lower than the emitter. The estimation error of DOA is the RMSE from the fit of the inverse tangent within direction finding range. The experimental results show that the system could determine the DOA of different vehicles in the presence of wind and the accuracy is within 6° in relative range and wind scale.

In

In general, the aperture of our system is very small (4 cm) which is an advantage for portability and mobility, but a challenge for high accuracy direction finding. Compared with other systems, our system design has a moderate sampling rate and computational complexity. In systems No. 1–3 (

In this paper, a real-time direction finding system is implemented based on a SHARC DSP processor. An approximation of the narrowband MUSIC algorithm is applied in the system for its advantages of accuracy and relatively low complexity for a small aperture array. By means of spatial coherence, the influence of wind noise is greatly reduced and the direction finding performance is enhanced. Experiments at different locations have demonstrated that the system is able to locate different types of vehicles with an accuracy of 6°. The system is mainly designed for vehicle direction finding using a UGS system. However, the system could also provide a reference for other applications such as video conferencing and speaker tracking.

The authors would like to thank the associate editor and anonymous reviewers for their valuable comments and suggestions to improve this paper.

The authors declare no conflict of interest.

Microphone arrays designed for verification of accuracy of direction finding. The MEMS microphone is ADMP504.

Block diagram of the MEMS microphone array system.

Photograph of the MEMS microphone array system, array aperture is 4 cm.

User interface of real-time DOA by UART.

A schematic diagram of direction finding method.

(

Flowchart of the moving vehicle direction finding method and time complexity.

(

(

The root-mean-square error (RMSE) of direction finding using MUSIC algorithm for different microphone arrays.

| |||||||
---|---|---|---|---|---|---|---|

| |||||||

Four microphones | |||||||

0.5 | 8.33 | 4.05 | 1.89 | 35.12 | 13.06 | 5.94 | 16.75 |

2 | 2.01 | 1.03 | 0.50 | 6.16 | 3.23 | 1.57 | 3.25 |

4 | 1.03 | 0.50 | 0.25 | 3.16 | 1.53 | 0.80 | 2.25 |

Three microphones | |||||||

0.5 | 9.14 | 4.48 | 2.23 | 38.67 | 15.33 | 7.07 | 30.50 |

2 | 2.25 | 1.15 | 0.56 | 7.29 | 3.44 | 1.82 | 15.46 |

4 | 1.15 | 0.54 | 0.28 | 3.60 | 1.81 | 0.93 | 8.46 |

Computational complexity of the proposed method.

Time | 7.2 ms | 38.0 ms | |||

| |||||

FFT | Calculation of the covariance matrix | Calculation of eigenstructure |
Peak search | ||

Percentage of each part | 15.6% | 6.6% | 10.2% | 67.6% |

The covariance matrix in STEP-3 is a Hermitian Matrix and the method of Jacobi for the eigen-decomposition algorithm of Hermitian matrix is used in this method [

Experimental results.

^{2}) | ||||
---|---|---|---|---|

Electric bicycle | 3 | ≥20 | 5.82 | 5,000 |

Car | 4 | ≥40 | 3.10 | 10,000 |

Bus | 4 | ≥40 | 2.20 | 20,000 |

Truck | 4 | ≥40 | 2.61 | 20,000 |

Tracked vehicle | 5 | ≥40 | 2.74 | 50,000 |

Performance comparison with other systems.

1 | Car [ |
15 | 5 | ULA | 4 | TDE | 44.1 |

2 | Trailer [ |
20 | 20 | ULA | 3 | TDE | 48 |

3 | Motor vehicle [ |
102 | 12 | ULA | 7 | TDE | 10 |

4 | Tracked vehicle [ |
>100 ^{a} |
1.5 | UCA | 5 | IWM | 8.192 |

5 | Tank [ |
20.23 | <2 | UCA | 12 | CWM | N/A |

6 | AAV ^{b} |
N/A ^{c} |
High | Random | N/A | ML | 4.96 |

The aperture of the microphone array is not given, it is estimated from the picture in [

AAV means amphibious assault vehicle;

N/A means not available.

Time elapsed for different methods.

Time elapsed (s) | 0.0022 | 0.1020 | 0.0724 | 0.3114 | 0.0074 |