UAV-BASED MULTI-SENSOR SYSTEMS FOR SURVIVOR DETECTION UNDER DISASTER RUBBLE: METHODOLOGY AND PERFORMANCE ANALYSIS

​ABSTRACT

Locating survivors trapped beneath collapsed structures is one of the most time-critical challenges in disaster response. This paper develops and evaluates a methodology for deploying multi-rotor unmanned aerial vehicles (UAVs) equipped with heterogeneous sensor payloads thermal infrared imaging, acoustic microphone arrays, and electrochemical CO₂ detectors to detect and localize living victims under earthquake and industrial disaster debris. A probabilistic detection model is derived as a function of flight altitude, sensor-to-target range, and sensor characteristics. A cooperative Voronoi-partitioned search algorithm for multi-UAV teams is formulated and validated via MATLAB/Simulink simulation. A three-UAV configuration achieves a combined detection probability of 0.87 within a 25-minute endurance window while reducing area coverage time by 66% relative to single-UAV operations. GAN-based acoustic noise suppression raises the signal-to-noise ratio by 10 12 dB, enabling acoustic detection at ranges previously inaccessible due to rotor interference.

Keywords: UAV, search and rescue, sensor fusion, thermal imaging, acoustic detection, CO₂ sensing, cooperative search algorithm, disaster robotics.

1. INTRODUCTION

Natural and industrial catastrophes continue to exact a devastating toll on human life. Earthquakes alone account for tens of thousands of fatalities each year, with survival odds declining precipitously as time elapses. Statistical studies show that the probability of extracting a living victim drops from roughly 90% in the first critical hours to below 5% once 72 hours have passed [1]. Speed is therefore the decisive factor between life and death.

Conventional search-and-rescue (SAR) approaches carry inherent limitations: manual debris inspection is slow and hazardous; trained search dogs tire rapidly; acoustic probing devices demand near-contact proximity [2]. Each method, applied in isolation, leaves significant portions of a collapsed structure unsearched within the critical survival window.

Unmanned aerial vehicles offer a fundamentally different operational paradigm. Their ability to penetrate structurally compromised zones inaccessible to rescue personnel, combined with the capacity to carry diverse sensor payloads, positions UAVs as a transformative tool in post-disaster operations [3]. This paper makes three contributions: (1) a closed-form probabilistic detection model encompassing thermal, acoustic, and CO₂ channels; (2) a cooperative multi-UAV search algorithm with guaranteed coverage properties; and (3) a comparative simulation study quantifying the performance advantage of multi-agent deployment.

2. MATERIALS AND METHODS

2.1 UAV Platform and Payload Architecture

The reference platform is a mini-class multi-rotor UAV with a maximum take-off weight of 5 7 kg. Three complementary sensor types are integrated: a thermal imaging camera (7 14 μm), a four-channel acoustic microphone array (50 Hz 8 kHz), and an electrochemical CO₂ sensor (0 5000 ppm). This combination exploits distinct physiological signatures of a living human body heat, vocalization, and respiratory CO₂ ensuring failure modes of the three channels are largely uncorrelated. Specifications are summarized in Table 1.

Table 1.

UAV and payload technical specifications

2.2 Probabilistic Detection Model

The probability of detecting a survivor is modeled as a function of the slant distance r and flight altitude h. For each sensor modality an exponential likelihood captures detection probability degradation with range:

where λᵢ is the attenuation coefficient for sensor i. Under the assumption that sensor failures are stochastically uncorrelated, the combined detection probability is:

For the thermal infrared channel, detection depends on the ratio of thermal contrast ΔT to the camera's noise-equivalent temperature difference , where is the standard normal and The acoustic channel links detection probability to the signal-to-noise ratio where α_AC ≈ 0.15 dB/m. CO₂ transport is modeled via the steady-state diffusion equation with wind advection, with human exhalation rate Q ≈ 200 mL/min and diffusion coefficient D ≈ 0.12 m²/s.

2.3 Cooperative Multi-UAV Search Algorithm

For a fleet of N UAVs assigned to a search zone of area S, spatial coverage is organized through Voronoi decomposition, partitioning S into N cells of approximately equal area. Within each cell, the assigned UAV executes a boustrophedon (lawnmower) trajectory. The inter-strip spacing d is constrained by the guaranteed detection radius of the thermal camera:

where P_th = 0.7 is the minimum required single-pass detection probability. The minimum coverage time is T_cover = S / (N v_UAV d). For N = 1, v_UAV = 10 m/s, S = 10,000 m², and d = 8 m, this yields T_cover = 125 s. Scaling to N = 3 reduces this to approximately 42 s a 66% reduction validated by simulation.

2.4 Rotor Noise Suppression via GAN-Based Filtering

Acoustic measurements aboard a hovering UAV are dominated by propeller tones that mask the low-amplitude signals produced by trapped survivors. An adaptive GAN filter is applied: ŝ(t) = G_θ[x(t), n_ref(t)], where x(t) is the raw microphone signal, n_ref(t) is a reference noise signal from the motor RPM sensor, and G_θ is the generator network. Training was performed on 500 noisy clean signal pairs with input SNR ranging from −5 to +10 dB [4].

3. RESULTS

3.1 Combined Detection Probability vs. Range

Numerical evaluation was performed for h = 15 m, T_env = 20 C, and T_body = 37 C. Results across the operationally relevant range of 2 20 m are presented in Table 2.

Table 2.

Per-sensor and combined detection probability as a function of range (h = 15 m)

At r = 10 m, the combined P_d^total = 0.906, compared to 0.64 for IR alone and 0.55 for acoustic detection. Sensor fusion delivers substantially higher detection probabilities than any individual channel, validating the complementarity hypothesis that motivated the multi-sensor architecture.

3.2 Multi-UAV Search Performance

Monte Carlo simulations in MATLAB/Simulink were conducted over a 100 × 100 m search zone with three randomly positioned survivors. One hundred independent trials were run per scenario. Results are presented in Table 3.

Table 3.

Search scenario comparison (S = 10,000 m², v = 10 m/s, 3 survivors)

The three-UAV configuration achieves P_detect = 0.87 within the 25-minute endurance window, compared to 0.71 for a single UAV. The expected number of survivors located rises from 2.1 to 2.6 out of three. Coverage time collapses from 125 s to 42 s, confirming the linear scaling predicted analytically.

3.3 Comparison with Conventional SAR Methods

Field data in the literature [1, 2] indicate that a five-person rescue team with trained search dogs requires 35 50 minutes to cover a 100 × 100 m zone, achieving detection probability 0.75 0.80. The three-UAV system achieves a superior detection probability (0.87) in approximately 3 4 minutes of active search a roughly tenfold speed advantage without exposing rescue personnel to the structural hazards of a freshly collapsed building.

4. DISCUSSION

The results make a strong case for sensor fusion as the architectural cornerstone of UAV-based SAR systems. No single sensor modality achieves P_d > 0.85 at ranges beyond 5 m. The thermal channel loses sensitivity as debris heats up; acoustic signals become inaudible under thick concrete layers; CO₂ disperses rapidly in wind. These failure modes are physically independent, which is why the joint reliability of the system remains high even when individual channels degrade.

GAN-based noise suppression deserves particular emphasis. Without filtering, rotor harmonics reduce effective acoustic SNR to levels at which P_d^AC at 5 m is approximately 0.15 essentially useless. After applying the GAN filter, SNR improves by 10 12 dB, elevating P_d^AC to 0.78. This is the difference between an acoustic channel that contributes meaningfully to the fusion product and one that adds noise to the joint estimate.

Several limitations constrain the generalizability of current findings. All results derive from simulation under idealized conditions. Real rubble environments are three-dimensional and dynamically hazardous: acoustic reflections distort source localization; fires generate spurious IR hotspots; wind produces complex CO₂ plume structures. Field validation at certified rubble-testing facilities is an indispensable next step before operational deployment. Within Kazakhstan's emergency response system, a three-UAV team can complete a preliminary survey of a zone up to 1 ha within 30 minutes, enabling targeted deployment of rescue teams rather than systematic grid search.

5. CONCLUSIONS

This paper has presented a methodology for UAV-based multi-sensor survivor detection under disaster debris. Three principal findings emerge:

The sensor fusion architecture achieves a combined detection probability of P_d^total = 0.906 at r = 10 m a 27-percentage-point improvement over thermal imaging alone (0.64), a difference that translates directly into survivors found or missed.

The cooperative Voronoi-partitioned search algorithm reduces zone coverage time by 66% when three UAVs operate in parallel, and pushes detection probability within the 25-minute endurance window to 0.87.

GAN-based acoustic noise suppression raises SNR by 10 12 dB, transforming the acoustic channel from a near-ineffective component (P_d^AC ≈ 0.15) into a substantial contributor to the fused detection probability (P_d^AC ≈ 0.78 at 5 m).

Future research priorities include field trials at dedicated rubble test sites, real-time integration with situational awareness systems, and development of three-dimensional survivor localization algorithms exploiting multi-UAV geometric diversity.

References:

1. Derekenaris G., Garofalakis J., Makris C. et al. Integrating GIS, GPS and GSM technologies for the effective management of ambulances // Computers, Environment and Urban Systems. 2001. Vol. 25, No. 3. P. 267 278.
2. Tomic T., Schmid K., Lutz P., et al. Toward a fully autonomous UAV: Research platform for indoor and outdoor urban search and rescue // IEEE Robotics & Automation Magazine. 2012. Vol. 19, No. 3. P. 46 56.
3. Baiocchi V., Dominici D., Mormile M. UAV Application in Post-Seismic Environment // ISPRS International Archives. 2012. Vol. XXXIX-B1. P. 87 92.
4. GAN Based Audio Noise Suppression for Victim Detection at Disaster Sites With UAV // IEEE Sensors Journal. 2023.
5. Murphy R.R. Disaster Robotics. Cambridge, MA: MIT Press, 2014. 432 p.
6. Marconi L., Melchiorri C., Beetz M., et al. The SHERPA Project: Smart collaboration between humans and ground-aerial robots for improving rescuing activities in alpine environments // Safety, Security, and Rescue Robotics. 2012. P. 1 4.