TL;DR
This paper explores the feasibility of using a drone-mounted MEMS microphone array to locate people in enclosed, low-visibility environments by distinguishing distress calls from the drone's own propeller noise. The research demonstrates that by filtering out low-frequency drone noise, distress signals can be localized with reasonable accuracy, paving the way for autonomous search and rescue applications.
What Problem Does This Paper Address?
Detecting and locating people in environments with reduced visibility, such as those affected by fire or smoke, is a critical challenge for search and rescue operations. Traditional visual sensors (cameras) become ineffective in such conditions. While device-based solutions exist, they rely on individuals carrying specific equipment, which isn't guaranteed. This paper tackles the problem of using sensor-based autonomous systems, specifically drones equipped with acoustic arrays, to find people in these challenging scenarios. The core difficulty lies in distinguishing faint distress signals (like cries for help) from the significant noise generated by the drone's propellers, which can easily mask the desired audio cues.
Methods
The study employs a drone equipped with a custom-built Uniform Planar Array (UPA) of 64 MEMS microphones. This array is designed to capture acoustic signals in the 4-16 kHz range. The system utilizes a deterministic wideband beamforming algorithm for Direction of Arrival (DOA) estimation. The processing pipeline involves:
- MEMS Acquisition: High-speed signal acquisition from each microphone.
- Signal Processing: Deinterlacing, decimation, and crucially, a high-pass filter to remove low-frequency noise (below 2 kHz) attributed to drone propellers and other environmental sources.
- DOA Assessment: Using wideband beamforming to estimate the angles of arrival (azimuth and elevation) of sound sources, creating an "acoustic image."
The researchers also developed a method to estimate a sound source's 3D position by triangulating the angles of arrival of the direct signal and its reflections from known surfaces (floor, ceiling, walls), assuming specular reflection.
Datasets
Not discussed in the paper. The experiments were conducted using controlled setups with generated acoustic signals (sinusoidal pulses and voice recordings) and controlled background noise levels simulating drone propeller noise.
Key Insights and Ideas
- Propeller Noise as a Discriminator: The key insight is that drone propeller noise is predominantly low-frequency. This characteristic allows for effective discrimination from higher-frequency distress signals using simple high-pass filtering.
- Passive Localization: The system enables passive localization by analyzing direct and reflected sound paths, eliminating the need for active sonar-like pulse transmission.
- Reflection-Based Triangulation: Utilizing reflections from enclosure surfaces, combined with the drone's known position, allows for 3D localization without direct range estimation.
- Trade-off with Noise: While the system performs well in low to medium noise conditions, high propeller noise levels can significantly degrade accuracy, highlighting the importance of acoustic isolation and stable flight.
Key Results
The experiments were conducted in three setups to evaluate the system's performance under varying conditions:
- Setup 1 (Noise Characterization): Characterized drone propeller noise, finding it mostly within 0-2 kHz.
- Setup 2 (Direct Signal Localization): Tested localization of a 5-kHz sinusoidal pulse. In low to medium noise (SNR 15 dB and 0 dB), the system achieved good accuracy with a mean azimuth error of -6° and elevation error of 0°. However, at high noise levels (SNR -15 dB), the errors increased significantly (azimuth -8° ± 1.9°, elevation -2° ± 1.2°).
- Setup 3 (Reflected Signals & Voice): Evaluated localization with two reflective surfaces and a voice signal.
- With a pulsed signal, accuracy remained good in low/medium noise (e.g., vertical plane azimuth: 0° ± 0.1°), but degraded with high noise (e.g., vertical plane azimuth: -2° ± 1.3°).
- With a voice signal, the accuracy was moderate, showing a bias of up to 4° and deviations up to 12° in high noise conditions.
Overall, the system demonstrated the feasibility of estimating DOA for both direct and reflected signals, with performance degrading significantly only at very high noise levels that could potentially saturate the MEMS sensors.
Limitations and Future Work
Limitations:
- High Noise Degradation: The system's accuracy significantly decreases in high propeller noise environments, potentially leading to saturation of MEMS sensors.
- Simplified Environments: The experiments were conducted in controlled setups with simplified geometries (two reflective surfaces), which may not fully represent complex real-world indoor environments with numerous non-perpendicular reflective surfaces.
- Drone Stability: The accuracy is highly dependent on the drone's stability; jerky movements lead to larger errors.
Future Work:
- Acoustic Isolation: Mechanically and acoustically isolating the MEMS array from the drone structure is recommended.
- Hovering for Estimation: Periodically hovering the drone can improve estimation accuracy by minimizing noise.
- Integration with New Drones: Adapting frequency parameters for different drone models.
- Advanced Integration: Integrating a 10x10 cm array with a 2D gimbal onto a new drone for real-world testing.
- Controller Integration: Sending acoustic source position information directly to the drone controller for autonomous navigation towards the detected person.
Summary updated