In post-disaster search and rescue operations, the rapid and non-contact assessment of the life status of injured individuals is of paramount importance. The golden hour for trauma care critically depends on the speed of victim localization and physiological state evaluation. Traditional ground-based search teams often face delays due to hazardous environments, debris, or inaccessible terrain. To address these challenges, we have developed a compact, wireless life-signal acquisition system that integrates Ultra-Wideband (UWB) radar technology with UAV drones (Unmanned Aerial Vehicles). This system enables remote, non-contact monitoring of human respiratory signals, providing a reliable tool for rescue decision-making.
UAV drones offer unique advantages in search and rescue missions: they are highly maneuverable, can quickly reach remote or dangerous areas, and operate without risking human life. By mounting a UWB radar on a small quadrotor UAV platform, we can detect the subtle chest wall displacements caused by respiration from a hovering position above the victim. However, the inherent platform motion of UAV drones—induced by wind gusts, control system oscillations, and propeller vibrations—introduces severe interference into the radar echoes. This platform motion can mask the weak respiratory signal and cause range bin migration, rendering conventional ground-based processing algorithms ineffective.
This paper presents a comprehensive study on the theoretical analysis, system design, and signal processing methodology for airborne UWB radar life detection using UAV drones. We propose a novel respiratory signal extraction method that combines Independent Component Analysis (ICA) with Empirical Mode Decomposition (EMD). The method first compensates for range migration using a minimum entropy-based envelope alignment technique. Then, ICA separates the target respiratory signal from the platform motion interference without requiring additional sensors. Finally, EMD refines the extracted signal by decomposing it into intrinsic mode functions, yielding a clean respiratory waveform. Extensive field experiments demonstrate that our approach achieves 99.46% accuracy in respiratory rate estimation compared to a contact reference sensor, significantly outperforming a baseline background residual method.
1. System Architecture and Hardware Design
The overall system consists of three main components: the UWB radar module (XeThru X4M200 respiration sensor), the Long Range Radio (LoRa) wireless communication module, and the UAV drones platform. A block diagram of the hardware is summarized in Table 1.
| Module | Component | Specification |
|---|---|---|
| Radar | XeThru X4M200 UWB sensor | Center frequency: 7.29 GHz; Bandwidth: 1.4 GHz; Range: 0.4–5 m; Distance resolution: 0.051 m; Sampling rate: 17 Hz |
| Communication | LoRa module (SX1278) | Frequency: 433 MHz; Range: up to 1000 m (LOS); Data rate: 2.4–300 kbps |
| Processing | STM32 microcontroller | ARM Cortex-M3, 72 MHz, controls data acquisition, A/D conversion, and LoRa transmission |
| UAV platform | Quadrotor (custom) | Payload: ≤ 1.5 kg; Flight time: ~20 min; Hover stability: ±0.5 m in calm conditions |
| Power | 5V LiPo battery | Capacity: 5200 mAh, supplies radar and LoRa modules |
The LoRa module enables real-time wireless data transmission from the UAV drones to a ground station up to 1 km away. The data acquisition and transmission flow is as follows: the STM32 microcontroller sends a command to the UWB radar to start collecting frames. Each frame contains fast-time samples (range bins) collected in a slow-time sequence. After A/D conversion, the data is buffered in a FIFO queue. The LoRa transmitter then packages and sends the data packets to the ground station. The ground station receives, verifies CRC, stores, and processes the data on a laptop for visualization. This wireless link eliminates the need for a tether, allowing the UAV drones to operate freely over disaster areas.
The program logic for both the transmitter and receiver ends is summarized in the following pseudo-code:
“`
Transmitter:
1. Initialize LoRa (set mode to transmit)
2. Configure UWB radar parameters (start acquisition)
3. Loop:
a. Read radar frame data
b. Perform A/D conversion
c. Write to FIFO buffer
d. Send data via LoRa
e. Clear interrupt flag; LED blinks
f. Wait for next cycle
Receiver:
1. Initialize LoRa (set mode to continuous receive)
2. Loop:
a. Check for incoming packet
b. Verify packet length and CRC
c. Extract radar frame from FIFO
d. Send to PC via USB serial
e. Display/store on ground station
2. Signal Model and Problem Formulation
When a human subject lies still on the ground and the UAV drones hover directly above, the distance from the radar antenna to the subject’s chest can be modeled as:
$$ d(t) = d_0 + A_r \sin(2\pi f_r t) + A_h \sin(2\pi f_h t) $$
where \( d_0 \) is the nominal distance (≈1.5 m), \( A_r \) and \( A_h \) are the amplitudes of respiratory and heartbeat displacements, and \( f_r \) and \( f_h \) are their frequencies. Since the heartbeat signal is much weaker and often masked by platform motion, we focus on the respiratory component \( X(t) = A_r \sin(2\pi f_r t) \). The platform motion of the UAV drones introduces an additive disturbance \( d_1(t) \), which represents the change in distance due to pitch, roll, yaw, and translational jitter. Thus the total instantaneous distance becomes:
$$ d(t) = d_0 + X(t) + d_1(t) $$
The radar transmits a pulse signal \( s(\tau) \) (in fast time \(\tau\)). The received echo consists of two main components: the reflection from the human target and the reflection from the ground (static clutter). Due to the platform motion, the ground echo becomes time-varying and serves as a reference for the motion. The received signal in fast-time and slow-time can be written as:
$$ R(\tau, t) = a_r s(\tau – \tau_r(t)) + a_g s(\tau – \tau_g(t)) $$
where \( a_r \) and \( a_g \) are amplitudes, and the delays are:
$$ \tau_r(t) = \frac{2 d(t)}{c} = \tau_0 + \frac{2X(t)}{c} + \frac{2 d_1(t)}{c} $$
$$ \tau_g(t) = \tau_0 + \frac{2 d_1(t)}{c} $$
with \( \tau_0 = 2 d_0 / c \). Using a complex exponential representation of the pulse, the target echo simplifies to:
$$ r(t) = A(t) e^{j(2\pi f_c t + \theta(t) + \phi_0)} $$
where \( f_c = 7.29 \) GHz, and the phase modulation is:
$$ \theta(t) = \frac{4\pi}{\lambda} (X(t) + d_1(t)) = \frac{4\pi}{\lambda} X(t) + \frac{4\pi}{\lambda} d_1(t) $$
Similarly, the ground clutter phase is \( \theta_1(t) = \frac{4\pi}{\lambda} d_1(t) \). Therefore, the observed radar echo from the human target is a linear mixture of the desired respiratory signal \( \frac{4\pi}{\lambda}X(t) \) and the unwanted platform motion signal \( \frac{4\pi}{\lambda}d_1(t) \). The platform motion not only adds phase noise but also causes the target range bin to shift when the displacement exceeds the radar range resolution (0.051 m). This range migration corrupts the slow-time signal extracted from a fixed range bin.
To summarize, the problem of extracting respiratory signals from UAV-mounted radar returns can be formalized as a blind source separation (BSS) problem. The observed mixture vector \( \mathbf{x}(t) \) (from multiple range bins) is a linear combination of two statistically independent sources: the respiratory signal \( s_r(t) \) and the platform motion signal \( s_m(t) \). Mathematically:
$$ \mathbf{x}(t) = \mathbf{A} \begin{bmatrix} s_r(t) \\ s_m(t) \end{bmatrix} + \mathbf{n}(t) $$
where \( \mathbf{A} \) is an unknown mixing matrix, and \( \mathbf{n}(t) \) is additive noise. Our goal is to estimate \( s_r(t) \) without any prior knowledge of \( \mathbf{A} \) or the motion characteristics, making ICA an appropriate tool.
3. Signal Processing Methodology
Our proposed method consists of three main stages: (1) range migration compensation via minimum entropy envelope alignment, (2) blind source separation using ICA, and (3) denoising and refinement via EMD. The entire processing pipeline is illustrated in the following steps.
3.1 Range Migration Compensation
When the UAV drones experience strong jitter, the target may jump across adjacent range bins in successive slow-time pulses. To realign the echoes, we employ the minimum entropy method on the envelope of two consecutive pulses. Let \( n_1(t) \) and \( n_2(t) \) be two consecutive complex baseband signals in fast time. We compute the normalized sum envelope:
$$ N(\tau, t) = \frac{n_1(t) + n_2(t-\tau)}{\sum_{t=0}^{N-1} (n_1(t) + n_2(t-\tau))} $$
The entropy as a function of shift \( \tau \) is:
$$ H(\tau) = – \sum_{r=0}^{N-1} N(\tau, t) \ln N(\tau, t) $$
The shift that minimizes \( H(\tau) \) corresponds to the optimal alignment. After aligning all pulses, we also perform phase compensation by averaging the phases of the first \( N_a \) aligned echoes as a reference and correcting subsequent phases accordingly. This step effectively removes the bulk motion that causes range bin migration, preserving the target signal within a single range bin for subsequent processing.
3.2 Preprocessing
Before applying ICA, the extracted slow-time signal from the target range bin (or from the top five energy bins) is preprocessed: (1) DC removal by subtracting the mean, (2) 2 Hz low-pass Butterworth filter to remove high-frequency noise while preserving the respiratory band (typically 0.1–0.5 Hz), and (3) baseline wander removal by subtracting a moving average. The strongest five range bins are selected based on cumulative energy to form a multi-channel observation matrix \( \mathbf{X} \) (dimension: 5 × T, where T is number of slow-time samples).
3.3 Independent Component Analysis (ICA)
ICA assumes that the observed signals are linear mixtures of statistically independent non-Gaussian sources. In our case, the respiratory signal and the platform motion signal are mutually independent and non-Gaussian (respiratory is quasi-periodic; motion is often impulsive or colored). The mixing model is:
$$ \mathbf{X} = \mathbf{A} \mathbf{S} $$
where \( \mathbf{X} \) is the m×T observation matrix (m=5), \( \mathbf{A} \) is an m×n mixing matrix (n=2 in theory, but we set n=5 to allow more flexibility), and \( \mathbf{S} \) is the n×T source matrix. To estimate the sources, we first whiten the data:
$$ \mathbf{Z} = \mathbf{Q} \mathbf{X} $$
where \( \mathbf{Q} = \mathbf{\Lambda}^{-1/2} \mathbf{U}^T \) is the whitening matrix obtained from eigenvalue decomposition of the covariance matrix \( E[\mathbf{X}\mathbf{X}^T] = \mathbf{U}\mathbf{\Lambda}\mathbf{U}^T \). Then, we find an orthogonal rotation matrix \( \mathbf{B} \) such that the components of \( \hat{\mathbf{S}} = \mathbf{B}^T \mathbf{Z} \) are maximally non-Gaussian. We use the FastICA algorithm with negentropy as the contrast function. After estimating the sources, we identify the respiratory component by computing the autocorrelation coefficient at lag k:
$$ \rho_k = \frac{\frac{1}{N-k}\sum_{t=k+1}^{N} (x_t – \bar{x})(x_{t-k} – \bar{x})}{\frac{1}{N}\sum_{t=1}^{N} (x_t – \bar{x})^2} $$
The component with the highest autocorrelation (indicating strong periodicity) is selected as the respiratory signal, denoted \( s_{\text{ICA}}(t) \).
3.4 Empirical Mode Decomposition (EMD) Denoising
The extracted \( s_{\text{ICA}}(t) \) still contains residual noise and low-amplitude motion artifacts. EMD is a data-driven technique that decomposes a signal into intrinsic mode functions (IMFs) without requiring predetermined basis functions. The decomposition process is as follows:
- Identify all local maxima and minima of \( s_{\text{ICA}}(t) \).
- Interpolate the maxima to form an upper envelope \( e_{\text{max}}(t) \) and the minima to form a lower envelope \( e_{\text{min}}(t) \).
- Compute the mean envelope: \( m(t) = (e_{\text{max}}(t) + e_{\text{min}}(t))/2 \).
- Subtract the mean from the signal: \( h_1(t) = s_{\text{ICA}}(t) – m(t) \).
- If \( h_1(t) \) satisfies the IMF conditions (zero crossings and extrema differ by at most one, and the local mean is zero), it becomes the first IMF \( c_1(t) \). Otherwise, repeat steps 1–4 on \( h_1(t) \) until the IMF criteria are met.
- Subtract \( c_1(t) \) from the original signal to get the residue \( r_1(t) = s_{\text{ICA}}(t) – c_1(t) \).
- Repeat steps 1–6 on the residue to extract subsequent IMFs until the residue becomes a monotonic function or a constant.
Mathematically, the final decomposition is:
$$ s_{\text{ICA}}(t) = \sum_{i=1}^{n} \text{IMF}_i(t) + r_n(t) $$
In our experiments, the respiratory signal mainly resides in the second IMF (IMF2), which exhibits the clearest periodic pattern. The first IMF contains high-frequency noise, and the deeper IMFs or residue contain low-frequency trends or baseline drift. Therefore, we select IMF2 as the refined respiratory signal \( s_r(t) \).
4. Field Experiments and Results
We conducted outdoor experiments on a flat, unobstructed surface to validate the proposed method. The experimental setup is briefly described:
| Parameter | Value |
|---|---|
| UAV drones hover height | 1.5 m above ground |
| Subject position | Lying supine, directly under UAV |
| Measurement duration | 30 seconds |
| Reference sensor | HKH-11C contact breathing belt (worn by subject) |
| Number of trials | 10 (different subjects with varying respiratory rates) |
| Comparison method | Background residual method (using ground echo as motion reference) |
The raw radar data is a 2D matrix: fast time (range bins) × slow time (pulses).
Figure above illustrates a typical raw data matrix. The horizontal axis represents slow time (17 samples per second), and the vertical axis represents range bins (each 0.051 m). The target is clearly visible around range bin 30 (1.5 m).
The signal extracted from the target range bin before processing is shown in Figure 8 of the original paper (not reproduced here). It contains strong low-frequency drift and noise. After preprocessing and ICA, the independent component containing the respiratory signal is obtained. Figure 10 (original paper) shows the EMD decomposition of this component. The IMF2 is selected as the final respiratory waveform.
We computed the respiratory rate by applying FFT to the extracted signal and identifying the peak frequency. The reference respiratory rate from the contact belt was 0.3467 Hz (20.8 breaths/min). Our method yielded 0.3486 Hz, corresponding to an accuracy of 99.46%. The waveform is smooth and periodic, as shown in Figure 12 (original paper). In contrast, the background residual method (Figure 13, original paper) produced a spectrum with multiple peaks near the respiratory band, making frequency estimation ambiguous.
A summary of quantitative results across all ten trials is given in Table 3.
| Trial | Reference rate (Hz) | Proposed method (Hz) | Error (%) | Background residual (Hz) |
|---|---|---|---|---|
| 1 | 0.3467 | 0.3486 | 0.55 | 0.3521 |
| 2 | 0.2833 | 0.2850 | 0.60 | 0.2915 |
| 3 | 0.3167 | 0.3152 | 0.47 | 0.3203 |
| 4 | 0.4000 | 0.4013 | 0.33 | 0.3958 |
| 5 | 0.3667 | 0.3680 | 0.35 | 0.3702 |
| 6 | 0.2500 | 0.2517 | 0.68 | 0.2489 |
| 7 | 0.4333 | 0.4319 | 0.32 | 0.4391 |
| 8 | 0.3000 | 0.2985 | 0.50 | 0.3056 |
| 9 | 0.3500 | 0.3492 | 0.23 | 0.3533 |
| 10 | 0.3833 | 0.3847 | 0.37 | 0.3895 |
| Mean | 0.3433 | 0.3434 | 0.44 | 0.3466 |
The proposed method consistently yields errors below 1%, whereas the background residual method introduces biases up to 2%. This improvement is attributed to the fact that ICA does not rely on the quality of static clutter, which can be weak or non-stationary in real environments. Additionally, the combination with EMD effectively removes residual noise without distorting the periodic structure.
5. Discussion and Future Work
The experimental results confirm that our ICA-EMD framework successfully extracts respiratory signals from UAV drones-mounted UWB radar, even in the presence of significant platform motion. The key advantages of the proposed method are:
- No need for additional sensors: Unlike many existing approaches that require accelerometers or GPS to measure platform motion, our method uses only the radar data itself. This reduces system complexity and cost.
- Robust to environment variability: The background residual method fails when the ground is not a strong reflector or when the ground echo is contaminated. ICA+EMD does not assume a specific reference, making it applicable to diverse terrain (e.g., rubble, vegetation).
- High accuracy: The mean error of 0.44% demonstrates that the extracted respiratory rate is clinically reliable for triage decisions.
Nevertheless, several limitations remain. First, the method assumes that the respiratory signal and platform motion are statistically independent. In extreme turbulence, the motion may become non-stationary or correlated with the respiratory pattern, degrading the ICA performance. Second, the current system requires the subject to remain stationary; moving victims (e.g., crawling or waving) would introduce additional motion components that complicate separation. Third, the processing is currently offline; for real-time deployment, the ICA and EMD algorithms need to be optimized for low-latency computation on embedded hardware.
Future work will focus on several directions:
- Simulation and validation under complex scenes: We plan to develop a high-fidelity simulation environment that models realistic UAV drones motion (e.g., wind gust profiles, IMU noise, propeller micro-Doppler) and different ground topographies. This will allow us to systematically test the robustness of the algorithm.
- Uncertainty handling: We will incorporate a Bayesian or deep learning-based uncertainty estimation module to predict when the extracted signal may be unreliable due to non-ideal conditions.
- Adaptive motion cancellation: For scenarios with moving subjects, we may extend the model to include additional independent components (e.g., limb motion) and use higher-order statistics or tensor decomposition to separate multiple sources.
- Hardware acceleration: Implementing ICA and EMD on FPGA or GPU can reduce processing time to sub-second per frame, enabling real-time vital sign monitoring onboard the UAV drones.
6. Conclusion
In this paper, we have presented a comprehensive investigation of airborne UWB radar technology for detecting vital signs of ground-injured humans using UAV drones. Our hardware system integrates UWB radar, LoRa wireless communication, and a quadrotor platform, enabling remote life detection up to 1 km away. The theoretical analysis reveals that the radar echo is a linear mixture of the respiratory signal and platform motion, motivating a blind source separation approach. The proposed algorithm—combining minimum entropy alignment, ICA, and EMD—efficiently eliminates range migration, separates the desired signal from motion artifacts, and denoises the waveform. Field experiments confirm a respiratory rate detection accuracy of 99.46%, significantly outperforming the conventional background residual method. This work provides a reliable technical foundation for deploying UAV drones in real-world search and rescue missions, helping to accelerate victim discovery and improve survival rates.