PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning-Assisted UAV Localization Framework for Post-Disaster Search and Rescue Missions

Xiangjian Gao, Hamid R. Sadjadpour

arXiv:2609.02916v1eess.SPcs.NI

TL;DR

Post-disaster SAR teams need to locate victims when infrastructure is destroyed and only low-power smartphone beacons are available. The paper presents a cooperative UAV framework that classifies each link and activates environment-specific 3D localization solvers. Simulations report improved localization accuracy and reduced runtime, including under unknown transmit-power and path-loss conditions.

  • Problem

    Reliable localization is lacking for survivors emitting low-power, intermittent beacons across heterogeneous propagation environments without communication infrastructure.

  • Method

    A centralized cooperative UAV pipeline uses CNN–LSTM LOS/NLOS classification to select Taylor–WLS, iterative LSRE–SOCP, or projection-based localization methods.

  • Results

    The framework reduces average RMSE from 66.3 m to 47.7 m and runtime from 5.28 s to 3.03 s in end-to-end NYUSIM validation.

  • Takeaways & Limitations

    The unified framework supports 3D UAV localization across LOS, NLOS, and mixed propagation environments, with a publicly released NYUSIM-based dataset for further research.

Abstract

from arXiv · show

The precise locating of trapped victims is arguably the most challenging issue in SAR operations, particularly when infrastructure is destroyed and SAR teams only have low-power beacon signals from smartphones to search with. This paper presents a framework for centralized and cooperative UAV-based localization with deep learning-based channel classification and 3D environment-adaptive target estimation. A CNN-LSTM classifier is employed based on a dataset generated at 867.5 MHz with NYUSIM. This classifier labels the individual links of a UAV to a target. Then, these labels activate specific localization solvers: a first-order Taylor-expanded WLS method for LOS settings, an LSRE-SOCP method with iterative refinement for NLOS, and a hybrid projection-based scheme for mixed environments. Through exhaustive simulations, this paper shows that the proposed framework significantly reduces runtime and yields high localization accuracy even when the transmit power and path-loss conditions are unknown. The method can be used in diverse propagation scenarios, making it easy to deploy in reality. The whole NYUSIM-generated dataset is made publicly available to support future research on disaster-aware wireless localization.

I. INTRODUCTION

Post-disaster SAR localization is difficult because infrastructure may fail and survivors may emit only weak, intermittent smartphone beacons. The paper addresses this with a centralized cooperative UAV framework that classifies links and selects localization methods for heterogeneous propagation environments.

  • Motivation: Disasters can trap survivors in inaccessible areas while destroying communication infrastructure and limiting rescue resources.These conditions make conventional communication and response methods unreliable.
  • Research gap: Existing approaches are limited because vision fails in smoke or debris, trajectory planning depends on prior maps, and many methods assume restricted propagation conditions.AoA and RSS remain feasible with standard hardware, but NLOS propagation introduces multipath, attenuation, bias, and unknown transmit-power or path-loss parameters.
  • Motivation: Smartphones may provide only brief, low-power device-ID broadcasts, making weak-signal localization fundamentally different from conventional methods.Because smartphones remain near their owners, locating the device can indicate the survivor’s location.
  • Framework: The proposed framework combines cooperative UAV data collection, CNN–LSTM per-link LOS/NLOS classification, and environment-adaptive localization solvers.The pipeline integrates channel classification, solver selection, and multi-UAV data fusion for disaster-response constraints.
  • Localization methods: The system uses first-order Taylor-based WLS for LOS, label-aware projection with LSRE–SOCP for mixed links, and enhanced LSRE–SOCP for NLOS conditions.The NLOS formulation jointly estimates target position, transmit power, and path-loss exponent, while the mixed strategy uses a trusted LOS bearing.
  • Dataset: A NYUSIM-based dataset contains LOS and NLOS measurements for training and evaluating learning-assisted UAV SAR localization, and is publicly released.The dataset targets outdoor SAR evaluation with co-registered wireless measurements and channel labels.

A. Angle of Arrival Measurements

The framework models UAV measurements through noisy AOA and RSS observations, using distinct LOS and NLOS propagation models for 3D localization. LOS measurements use geometric angles and log-normal RSS, while NLOS signals include obstruction-related attenuation.

  • Angle of Arrival Measurements: Each UAV estimates the target’s azimuth and elevation using an antenna array, with measured angles corrupted by noise.The target location is represented in 3D, and the UAV coordinates are known.
  • RSS Modeling under LOS and NLOS Conditions: RSS provides a low-complexity distance-related measurement, averaged across M temporal samples for each UAV link.The received power samples are indexed by UAV and time sample.
  • RSS Modeling under LOS and NLOS Conditions: Under LOS conditions, received power follows a log-normal shadowing model involving transmit power, UAV–target distance, path-loss exponent, and Gaussian shadowing noise.The path-loss exponent may vary across UAV links because of geometry, elevation, and local environment.
  • RSS Modeling under LOS and NLOS Conditions: NLOS propagation adds obstruction-dependent attenuation whose bias varies with materials, obstacle geometry, and UAV–target orientation.The additional attenuation can dominate localization error in highly obstructed disaster environments.
  • Empirical Channel Modeling in Disaster Scenarios: The empirical NLOS model uses 900 MHz propagation measurements and separates total distance into free-space LOS and obstructed NLOS segments.The total distance is expressed as d_k = d_NLOS + d_LOS, with separate path-loss components.

III. CHANNEL IDENTIFICATION

The channel-identification framework classifies links as LOS or NLOS from CIR-derived inputs using both engineered-feature classifiers and raw-sequence deep learning models. It supports SVM, GPC, and RF baselines while motivating direct CIR learning for dynamic disaster environments.

  • III. CHANNEL IDENTIFICATION: Channel identification uses CIR measurements to distinguish LOS and NLOS links before selecting localization methods.The framework includes preprocessing, feature extraction, and classification stages.
  • Machine Learning-Based Classification: Statistical CIR features include mean, standard deviation, skewness, kurtosis, log-mean, Rician K-factor, and fading-model goodness-of-fit metrics.The normalized features are concatenated into an input vector p(i) ∈ R^d.
  • Machine Learning-Based Classification: SVM classifies engineered feature vectors with a kernel-based decision function, with inference complexity scaling as O(N_SV · N_f).N_SV is the number of support vectors and N_f is the feature dimension.
  • Machine Learning-Based Classification: GPC produces probabilistic predictions through a Gaussian-process model and probit link, but its training and inference complexities scale as O(N_s^3) and O(N_s^2).N_s denotes the number of training samples.
  • Machine Learning-Based Classification: RF combines independently trained decision trees through majority voting, with inference complexity scaling as O(N_t · H_tree).N_t is the number of trees and H_tree is their average depth.
  • Machine Learning-Based Classification: Engineered-feature classifiers depend on feature-distribution stability between training and deployment, motivating DL models that learn directly from raw CIR sequences.The paper identifies this stability assumption as fragile in dynamic post-disaster environments.

B. Deep Learning-Based Classification

The proposed CNN–LSTM classifier combines local CIR-pattern extraction with temporal modeling across packets to classify channel links. It learns representations from raw CIR magnitudes and outputs binary LOS/NLOS probabilities.

  • Deep Learning-Based Classification: The CNN–LSTM architecture learns discriminative representations from raw CIR magnitudes without manual feature design.This combines CNN spatial modeling with LSTM temporal modeling under nonstationary propagation.
  • Deep Learning-Based Classification: CNN filters extract localized CIR patterns such as fading dips or spectral notches that may indicate NLOS propagation.The convolution operates on each packet’s CIR magnitude vector.
  • Deep Learning-Based Classification: LSTM layers model temporal dependencies across packet-level CNN embeddings, capturing interpacket dynamics in the CIR sequence.The final hidden state aggregates temporal information after processing all packets.
  • Deep Learning-Based Classification: The classifier projects the final LSTM hidden state to a binary probability and trains it using binary cross-entropy loss.The output layer uses a sigmoid activation with learnable parameters.

IV. PROPOSED LOCALIZATION ALGORITHM

The localization pipeline selects an environment-specific solver after classifying the UAV–target links. It uses first-order linearized WLS for LOS measurements and covariance-aware estimation with favorable complexity and CRLB behavior under low noise.

  • PROPOSED LOCALIZATION ALGORITHM: Per-link LOS/NLOS labels activate one of three localization algorithms for purely LOS, NLOS, or mixed propagation environments.Each solver is matched to the measurement behavior associated with its propagation condition.
  • LOS Environment Localization Algorithm: The LOS solver estimates 3D target position by combining RSS ratios and AOA constraints in a cooperative WLS formulation.RSS ratios eliminate the unknown transmit power, while the LOS model assumes approximately equal path-loss exponents across links.
  • LOS Environment Localization Algorithm: First-order Taylor linearization models angle and RSS-ratio perturbations through an error vector assembled from azimuth, elevation, and RSS noise.The resulting Jacobian blocks capture how measurement noise affects the linearized constraints.
  • LOS Environment Localization Algorithm: The fusion center applies WLS using the propagated error covariance Q = C W C^T.W represents the covariance of the measurement-noise vector, while C maps measurement noise into constraint errors.
  • LOS Environment Localization Algorithm: O(N_meas) computational complexity makes the LOS estimator scale linearly with the number of RSS and AOA measurements.N_meas denotes the total number of available measurements.
  • LOS Environment Localization Algorithm: When measurement noise is sufficiently small, WLS performance approaches the Cramér–Rao lower bound.This statement applies to the low-noise regime described for the estimator.

B. NLOS Environment Localization Algorithm

The NLOS localization method uses robust LSRE optimization to handle bounded link biases, then replaces computationally heavy SDP lifting with a convex SOCP formulation and iterative refinement for unknown transmit power and path-loss parameters.

  • SDP formulation: The LSRE objective is transformed through variable substitution, auxiliary variables, and matrix lifting into a semidefinite program with a standard rank-1 relaxation.The lifted variable is Z = zz^T, and the resulting complexity is characterized as O(N_lift^3.5).
  • SOCP formulation: A rotated-cone SOCP reformulation preserves convexity while eliminating matrix lifting and substantially lowering computational cost.The formulation introduces auxiliary variables for scaled distance, reciprocal, and quadratic terms.
  • SOCP formulation: The SOCP solution recovers the target position from the optimized variables and has polynomial computational complexity in the number of optimization variables.The cited complexity is approximately O(N_opt^3).
  • Iterative refinement: Because transmit power and path-loss exponents are unknown and variable across links, iterative refinement jointly updates the target position, transmit power, and path-loss parameters.The cycle uses initialization, SOCP solving, distance recomputation, RSS-based transmit-power updates, and path-loss inversion until convergence or a maximum iteration limit.
  • Iterative refinement: A relaxation step size smooths transmit-power updates, while RSS-error monitoring and early termination help detect and mitigate divergence.The relaxation parameter satisfies ν ∈ (0, 1].

C. Mixed Environment Localization Algorithm

The mixed-environment solver combines one LOS direction with NLOS RSS constraints in a channel-aware LSRE–SOCP formulation, then conditionally projects the estimate onto the LOS ray.

  • Mixed-environment model: The mixed case assumes exactly one LOS UAV and uses its reliable AOA direction together with LOS and NLOS RSS observations.The remaining UAV links are treated as NLOS.
  • Mixed-environment model: The channel label vector identifies LOS and NLOS links, with UAV 1 designated as the unique LOS anchor.The labels are binary, with y_k = 1 for LOS and y_k = 0 for NLOS.
  • Mixed LSRE–SOCP: The mixed LSRE objective sets zero bias and fixed β_LOS for the LOS link while retaining per-link bias intervals for NLOS links.The resulting conic constraints are solved for an unconstrained three-dimensional target estimate.
  • Projection refinement: The unconstrained estimate is projected onto the LOS ray to enforce geometric consistency because one LOS direction can leave residual drift along that axis.The projected estimate is compared with the unconstrained estimate using RSS mismatch cost.
  • Projection refinement: If projection increases the RSS mismatch beyond the tolerance, the method rejects it and retains the underlying LSRE–SOCP estimate.This fallback addresses false-positive LOS classifications without forcing the incorrect geometric constraint.

V. DISCUSSION

The framework is organized as a channel-aware pipeline that classifies each UAV–target link, infers the propagation environment, selects a matching localization solver, and outputs a 3D target estimate. Runtime measurements use consumer-grade hardware, and the dataset is publicly available.

  • Evaluation and availability: Runtime and inference evaluations run on an Intel Core i7-11700 desktop with 16 GB of RAM, while the complete NYUSIM dataset is publicly available.The reported elapsed times come from Python and MATLAB implementations.
  • Framework pipeline: The framework collects RSS and AOA measurements from at least four UAVs before classifying each link with a CNN–LSTM model.The classifier assigns binary LOS/NLOS labels to individual UAV–target links.
  • Environment inference: The inferred environment is LOS when at least two links are LOS, mixed when exactly one is LOS, and NLOS when none are LOS.These cases determine the subsequent localization algorithm.
  • Solver selection: The selected solvers are Taylor-WLS for LOS, iterative LSRE–SOCP for NLOS, and LSRE–SOCP with conditional LOS-ray projection for mixed propagation.The mixed solver compares projected and unprojected RSS mismatch costs before choosing the estimate.
  • Output: The pipeline outputs an estimated three-dimensional target location.The output is denoted x̂.

A. Dataset Generation

The dataset uses NYUSIM urban-macrocell simulations at 867.5 MHz to generate LOS/NLOS-labeled wireless measurements with realistic 3D UAV configurations, mobility, blockage, and angle information.

  • Simulation setup: NYUSIM v4.0 generates 41,000 samples at 867.5 MHz with 3 MHz bandwidth under Urban Macrocell scenarios.The distance distribution is 50% within 100 m, 30% within 100–150 m, and 20% within 150–200 m.
  • Geometry and propagation: The coordinate transformation fixes the target at [0, 0, 0] while dynamically generating UAV positions with user heights from 0 to 20 m.Foliage and outdoor-to-indoor penetration losses are sampled according to the LOS/NLOS label.
  • Dynamic scenarios: The data include stationary and mobile scenarios through spatial consistency and random human blockage states intended to reflect disaster situations.The scenarios represent trapped victims and fleeing individuals in events such as earthquakes, wildfires, and floods.
  • Measurement features: Each sample provides 3D receiver positions, multipath parameters, and AOA information, with LOS angles recomputed geometrically and rounded to the nearest degree.NLOS angles are set to zero in the dataset construction.

B. Deep Learning Validation

The study evaluates channel classifiers under identical train–test conditions and selects CNN–LSTM for practical deployment despite RF’s higher accuracy on the current split. CNN–LSTM operates on raw CIR sequences, achieves the highest AUC, and avoids manual feature engineering.

  • Evaluation setup: An 80/20 train–test split evaluates accuracy, precision, recall, F1-score, AUC, and inference time under identical conditions.The test set computes standard performance metrics, while all models are evaluated fairly.
  • Traditional classifiers: RF achieves the highest test accuracy (0.9519) and precision (0.9448) among the representative traditional classifiers.The comparison covers RF, GPC, and SVM using handcrafted statistical features from normalized CIR magnitudes.
  • CNN–LSTM design: CNN–LSTM processes raw CIR magnitude sequences, with convolutional layers extracting localized spectral features and LSTM layers modeling temporal multipath dependencies.This removes the need for manually engineered statistical features.
  • CNN–LSTM results: CNN–LSTM achieves 0.9290 test accuracy, 0.9108 precision, 0.9502 recall, 0.9301 F1-score, and the highest AUC of 0.9696.Its architecture contains approximately 7.2×10^4 trainable parameters and requires about 2–3 MFLOPs per inference.
  • Deployment rationale: CNN–LSTM is favored for deployment because it operates directly on raw CIR data, whereas RF’s handcrafted features are tied to the current NYUSIM setup.Porting RF to different regions, propagation statistics, or hardware would require additional data collection, retraining, and expert intervention.

C. Analytical Localization Algorithm Validation

The analytical validation tests environment-specific localization solvers across LOS, NLOS, and mixed conditions using three-dimensional Monte Carlo simulations. Results support Taylor-based LOS estimation, iterative SOCP for NLOS, and label-aware projection for mixed links, including substantial mixed-case gains.

  • Validation assumptions: The validation uses simplified analytical LOS/NLOS models with perfectly known link states to isolate localization algorithms.Measurements include AOA and RSS generated from specified channel models.
  • Simulation design: Localization performance is measured by RMSE over 5000 independent three-dimensional Monte Carlo trials while varying UAV count and coverage radius.The scenarios vary K from 4 to 7 at R0 = 100 m and vary R0 from 100 m to 400 m with K = 4.
  • LOS validation: Under LOS conditions, WLS–Taylor remains close to the CRLB and outperforms SR–WLS and SDP/SOCP across tested radii despite increased AOA noise.When AOA noise increases from 1° to 3°, RMSE rises particularly for large R0, but WLS–Taylor retains the reported advantage.
  • NLOS validation: In NLOS conditions, iterative refinement updates transmit power and path-loss exponent to mitigate model mismatch, while the proposed SOCP solver lowers complexity and runtime.LSRE-SDP has the lowest accuracy, whereas the proposed solver is motivated by time-critical SAR requirements.
  • Operational interpretation: For K = 4 UAVs, the average localization error is approximately 57 m, reducing the search area to a 50–60 m region within a 200 m × 200 m zone.The paper frames this accuracy as operationally useful before finer tools such as thermal imaging or through-the-wall radar are applied.
  • Mixed-environment validation: The projection strategy anchors estimation to a trusted LOS AOA ray while incorporating RSS measurements through LSRE–SOCP.At larger coverage radii, performance degrades slightly because projection error increases and AOA reliability decreases.
  • Mixed-environment validation: The label-aware projection method reduces average RMSE from 56.7 m to 28.6 m and runtime from 2.87 s to 1.51 s under K = 4 and R0 = 100 m.Relative to Iterative RWLS–SDP, it further decreases RMSE by roughly 52% and runtime by about 70%.
  • Robustness: When LOS bearings become unreliable, a consistency test rejects harmful projections and returns to the robust LSRE–SOCP baseline.The authors identify LOS/NLOS classification correctness as the dominant mixed-environment factor in realistic deployment.

D. NYUSIM-Based End-to-End Environment-Aware Localization Validation

NYUSIM-based end-to-end simulations evaluate the complete environment-aware localization framework under realistic propagation conditions. The proposed method improves localization accuracy and remains computationally efficient, while its evaluation includes best-case AoA assumptions and CNN–LSTM-generated channel labels.

  • End-to-end framework: The CNN–LSTM classifies each link from CIR sequences, and the resulting LOS/NLOS labels and RSS measurements drive environment-adaptive localization.The framework uses the inferred propagation pattern to select the appropriate solver for the localization environment.
  • Evaluation assumptions: The NYUSIM experiments use optimistic AoA errors of approximately 1°, so larger angular errors would increase absolute RMSE without changing the qualitative ranking.The passage identifies this setting as a best-case upper bound on AoA quality, particularly where RSS dominates in NLOS and mixed environments.
  • Validation results: Across 5000 trials, the proposed framework reduces average RMSE from 66.3 m to 47.7 m and runtime from 5.28 s to 3.03 s.Its CDF rises more sharply and saturates earlier than those of three strong baselines, indicating tighter error bounds across heterogeneous conditions.
  • Baseline behavior: RWLS-SDP and LSRE-SDP can produce outliers exceeding 500 m without proper initialization, whereas the SOCP baseline is more stable but slower.The extreme cases were excluded from evaluation to maintain fairness.
  • Scope and efficiency: Runtime reduction is smaller under NYUSIM than under analytical channel models because fixed β = 2.5 provides only modest acceleration in mixed environments.The framework nevertheless remains described as efficient and scalable, while hardware-specific field calibration is outside the work’s scope.
  • Practical deployment: CNN–LSTM labels are imperfect and may be affected by signal ambiguity or noise, but confidence filtering or rerunning can verify abnormal results at low computational cost.The authors state that this overhead remains negligible relative to manual SAR missions.
Loading 2609.02916v1…