Aerial Acoustics for Human Localization

• Aerial acoustics is defined as using elevated microphone arrays to capture sound for 3D human localization through angular measurements.
• The approach employs sparse array geometries, robust calibration, and subspace algorithms like MUSIC to achieve sub-degree and sub-decimeter accuracy.
• Systems leverage code-multiplexed identification and efficient beamforming to enable concurrent tracking of multiple targets with minimal hardware.

Aerial acoustics for human localization encompasses the use of microphone arrays deployed in elevated or ceiling-mounted positions to determine the three-dimensional positions and identities of humans (or wearable tags) within indoor environments by angular acoustic measurements. Central to recent advances is angle-only localization, wherein direction-of-arrival (DoA) estimation from aerial arrays allows precise multi-target tracking with low infrastructure and hardware complexity relative to time-of-flight (ToF) or time-difference-of-arrival (TDoA) methods. Key systems exploit sparse microphone array geometries, robust calibration, code-multiplexed tags, and subspace algorithms to achieve sub-decimeter and sub-degree accuracy under challenging real-world acoustic conditions (Fischer et al., 16 Aug 2025, Fischer et al., 2024).

1. Sparse Microphone Array Geometry and Virtual Sensing

Aerial acoustic localization relies on compact arrays of MEMS microphones arranged in geometric patterns that maximize angular resolution and source enumeration capability. Uniform Rectangular Arrays (URA), such as an 8 × 8 grid (d ≈ 8.255 mm, spatial Nyquist ≈ 20.8 kHz), provide a baseline, but sparse geometries (e.g., Nested, Open-Box, Billboard) reduce sensor count while leveraging difference co-arrays to synthesize “virtual” sensors (Fischer et al., 16 Aug 2025, Fischer et al., 2024).

In these constructions, the set of virtual sensors is given by the hole-free co-array:

$\Delta = \{p_i - p_j\,|\,p_i, p_j \in S\}$

where $S$ is the set of physical sensor positions. The co-array—particularly when processed via spatial smoothing—enables estimation of more concurrent sources than there are physical sensors, a property central to scalable angle-based localization frameworks.

Array Geometry	Description	Experiments/Capabilities
URA 8×8	Full sensor grid	Baseline, 1.2° mean error, up to 14 sources
Nested	Dense core + sparse ring	Large virtual aperture, robust, multi-source
Open-Box	U-shaped arms on edge	Fine resolution, reduced sensors, 2–3 sources
Billboard	Centered row/column	Wide aperture, moderate cost
Random	Pathological baseline	Degraded performance

Controlled masking of microphones in software allows emulation of these geometries, enabling studies of trade-offs in hardware and array performance (Fischer et al., 2024).

2. Acoustic Sampling, Angular Resolution, and SNR Constraints

The maximum operational frequency of the array is limited by the spatial Nyquist criterion:

$f_N = \frac{c}{2d}$

where $c \approx 343\,\mathrm{m/s}$ is the speed of sound and $d$ is the inter-microphone spacing. For $d \approx 8.255\,\mathrm{mm}$, $f_N \approx 20.8\,\mathrm{kHz}$ (Fischer et al., 2024).

Angular resolution, for a far-field beamformer, scales as:

$$\Delta \theta \approx \frac{\lambda}{D} \approx \frac{c}{f D}$$

where $D$ is the array aperture. Subspace methods (e.g., MUSIC) can surpass this classical bound, thanks to the exploitation of the full covariance structure.

Experimental measurements indicate that sub-degree bearing accuracy is achievable: the URA reported 1.2° mean error below 70° elevation; Billboard, Nested, and Open-Box produce 1.35–1.5°; Random geometry yields >2° except at very low elevations (Fischer et al., 2024). Degradation occurs at high elevations as effective aperture contracts.

Signal-to-noise ratio (SNR) fundamentally bounds accuracy. In simulation, the URA attains 1.1° mean error at SNR=20 dB, rising to 8.0° at SNR=0 dB. Sparse arrays, particularly with spatial smoothing, maintain competitive performance except at extreme low SNR or for large numbers of sources (Fischer et al., 16 Aug 2025).

3. Direction-of-Arrival Estimation and Source Enumeration

DoA estimation adopts a narrowband model: for an array of $M$ microphones,

$$\mathbf{x} = \sum_{i=1}^K A_i\,\mathbf{v}_S(\bar \theta_i, \bar \phi_i) + \mathbf{n}$$

with $$\mathbf{v}_S(\bar \theta, \bar \phi)_n = \exp(j 2\pi (m_{x,n} \bar\theta + m_{y,n} \bar\phi))$$. Subspace algorithms, notably MUSIC, are applied to the covariance or, more generally, a spatially smoothed virtual co-array covariance $R_{ss}$:

$$R_{ss} = \frac{1}{L_x L_y}\sum_{p=0}^{L_x-1}\sum_{q=0}^{L_y-1} J_{p,q} R_{co} J_{p,q}^H$$

Larger virtual co-arrays allow resolution of up to $O(K^2)$ sources for $K$ physical sensors, conditional on array geometry and algorithmic regularization (Fischer et al., 16 Aug 2025, Fischer et al., 2024).

4. Code-Multiplexed Source Identification and Beamforming

To enable concurrent human localization and identity discrimination, massive-angle acoustic systems employ mutually orthogonal or nearly orthogonal codes—in particular, Zadoff–Chu (ZC) and spectrally balanced variants.

• Standard ZC: For prime $N$ and root $q$:

$$x[n] = \exp\left(-j\pi q n(n+1)/N\right),\quad n=0\ldots N-1$$

With ideal cyclic auto-correlation, but possible frequency tilt biasing DoA.

• Spectrally balanced ZC (SC-ZC, MS-ZC): These linear or Hermitian-symmetric combinations reduce estimation variance at high elevations by flattening the spectrum, at a moderate cost in autocorrelation peak amplitude ($\approx N/2$) (Fischer et al., 16 Aug 2025).

Post-DoA estimation, delay-and-sum beamformers extract the estimated $K$ sources. Assignment of each spatially separated code to source identity is performed through maximizing global confidence via correlation and the Hungarian algorithm.

Sequence Type	Peak Correlation	Identity Robustness
ZC	$N$	SNR-sensitive, frequency tilt
SC-ZC	$N/2$	Lower variance at high elevation
MS-ZC	$N/2$, Hermitian	Robust at varied angles

Identification robustness exceeds 95% for inter-source separations above 5° at SNRs as low as −20 dB (Fischer et al., 16 Aug 2025).

5. Self-Calibration and Deployment Considerations

A perspective-n-point (PnP) self-calibration protocol registers the spatial relation between arrays using just a moving tag, reducing deployment effort. For two arrays $A_1,A_2$ with unknown pose, observed DoA unit vectors and path-length variables are iteratively adjusted under known inter-array spacing until geometric consistency is achieved:

$$p_1 + R_1 d_1(t)\lambda_1(t) = p_2 + R_2 d_2(t)\lambda_2(t)$$

A nonlinear least-squares solution over all source instances yields the global registration without external tracking. This approach obviates per-node six-parameter manual surveying in favor of a small set of tracked trajectories (Fischer et al., 16 Aug 2025).

6. Experimental Performance and Practical Limitations

MASSLOC, an aerial acoustic framework, empirically achieves median 3D localization errors of 55.7 mm (50th percentile) for a single moving tag in a reverberant lobby ($T_{60}=1.6\,\mathrm{s}$, velocity up to 1.9 m/s), with angular errors of 0.84° (50th percentile) across two arrays. Identification of up to 14 simultaneous sources is error-free in anechoic conditions; three concurrent tags in reverberant settings localize with median 54 mm error, 95th percentile 178 mm, with performance degrading only when angular separations fall below 5° (Fischer et al., 16 Aug 2025).

Sparse geometries such as Nested, Billboard, Open-Box retain high performance (median ≈85–90 mm); Coprime, Random, or more heavily reduced-arrays suffer degraded resolution. In direct multi-source experiments, a full 8×8 URA array achieves 1.8° mean angular error for three simultaneous sources, while sparse geometries exhibit somewhat higher errors but competitive resolution for modest source counts (Fischer et al., 2024).

7. Towards Human Wearable Tag Localization and Future Prospects

The translation of aerial acoustics from active tags to direct human tracking necessitates attention to several factors. Human speech, extending from 100 Hz to 8 kHz, may require reconfigured microphone spacings (d ≈ 4 cm for $f_N≈4\,\mathrm{kHz}$), or employ ultrasonic tags (10–20 kHz) that modulate speech or emit codes. Foreseen challenges for wearable-tag localization include SNR reduction due to clothing/body absorption, non-stationary tag orientation (inducing Doppler and acoustic shadowing), and real-time data association for continuous trajectory estimation. Mitigation comprises power/SNR tuning, Doppler-robust codes, frequent angle estimation, and employing spherical stochastic filters such as the von Mises–Fisher Kalman filter.

Operationally, arrival angle outputs can achieve sub-decimeter and sub-degree tracking for tens of simultaneous subjects with only two or three aerial arrays, using <10 ms latency MUSIC and beamforming pipelines on modern FPGA platforms (Fischer et al., 16 Aug 2025). Hardware costs for MEMS-based URA arrays remain low (≈$2 per microphone,$150 for FPGA board). Practical deployment requires only modest calibration and array synchronization effort, with minimal installation compared to ToF/TDoA anchor networks (Fischer et al., 2024).

Aerial acoustics for human localization, as demonstrated by the MASSLOC pipeline and allied array processing research, enables robust, scalable, and cost-effective multi-target indoor localization suitable for applications ranging from collaborative robotics and asset tracking to wearable human subject localization under demanding acoustic conditions (Fischer et al., 16 Aug 2025, Fischer et al., 2024).