Abstract
This paper addresses the target localization problem in complex multipath propagation environments for three-dimensional (3-D) radar systems. We propose an approach based on the singular value decomposition (SVD) technique to reduce data dimensionality and formulate a joint multiple snapshot sparse representation problem in the signal subspace domain. A novel sparse representation-based Direction of Arrival (DOA) estimation algorithm is introduced, combining alternatingly iterative and dictionary refinement techniques. Theoretical benchmarks are established via closed-form Cramér-Rao bounds (CRB) for both target DOA and attenuation coefficient estimations. Experimental results from both simulated and measured data demonstrate that the proposed algorithm significantly enhances target localization accuracy in complex terrain and limited snapshot scenarios.
Introduction: The Challenge of Low-Angle Tracking
Radar tracking of low-altitude targets is a critical yet challenging task in modern surveillance. The primary obstacle is the "multipath" effect, where the radar receives not only the direct signal from the target but also reflected signals from the ground surface. When a target is at a low elevation angle, these signals are highly correlated and often arrive within the same radar beamwidth. This proximity makes them nearly indistinguishable for traditional signal processing methods.
Conventional techniques like MUSIC (Multiple Signal Classification) often rely on "spatial smoothing" to decorrelate these signals. However, this approach reduces the effective aperture size of the radar array, which in turn degrades the angular resolution. Furthermore, most existing models assume a perfectly flat, smooth reflecting surface. In reality, terrain is often complex, featuring hills, slopes, and varying vegetation. These deterministic models fail to account for such irregularities, leading to substantial errors in target altitude measurement.
Proposed Methodology: Sparse Representation and Dictionary Refinement
To overcome these limitations, we introduce a robust three-step framework based on Sparse Representation (SR). The core idea is to leverage the inherent sparsity of the target signal in the elevation domain.
1. SVD-based Signal Subspace Extraction
The process begins with Singular Value Decomposition (SVD) applied to multi-snapshot data. By projecting the received signal onto the signal subspace, we reduce the dimensionality of the problem and effectively increase the Signal-to-Noise Ratio (SNR). This preprocessing step ensures that the subsequent optimization is focused on the most relevant components of the signal, improving stability in low-SNR environments.
2. Composite Parametric Dictionary Construction
A key innovation of this work is the construction of a composite parametric dictionary. Unlike standard dictionaries that use simple steering vectors, our "atoms" represent the combined structure of the direct path and its multipath reflection. Each atom is a function of the target's DOA, the reflection's DOA, the complex attenuation coefficient of the ground, and the path difference between the direct and reflected signals. By embedding the physical relationship between the direct and reflected signals directly into the dictionary atoms, we reduce the problem's sparsity to one, simplifying the optimization process.
3. Three-Stage Adaptive Refinement
Searching a four-dimensional grid for the optimal parameters is computationally prohibitive. To address this, we employ an alternatingly iterative scheme:
- Stage 1: Coarse DOA Estimation. We perform an initial search for the target and image directions using a simplified symmetric model.
- Stage 2: Multipath Parameter Estimation. We estimate the multipath attenuation and path differences based on the initial DOA estimates.
- Stage 3: Joint Refinement. The dictionary grid is adaptively "shrunk" around the target's estimated location. This iterative refinement allows the algorithm to home in on the precise location without the need for an exhaustive global search.
Experimental Validation
The algorithm was rigorously tested using both simulated data and real-world measurements.
Simulation Results
Using a 16-element Uniform Linear Array (ULA), we tested the algorithm across various scenarios, including ideal smooth surfaces and complex sloped terrains. The results showed that the proposed method's Root Mean Square Error (RMSE) closely approached the theoretical Cramér-Rao Bound (CRB) at SNRs above 10dB. Notably, the method outperformed standard maximum likelihood algorithms like RML and AP-ML, particularly in scenarios with very few data snapshots. While subspace-based methods often require dozens of snapshots to form a reliable covariance matrix, our SR-based approach remains stable with as few as 10 snapshots.
Real-World Performance in Hilly Terrain
Field tests were conducted using a 20-channel elevation radar system located in a hilly area featuring groves and cropland. Ground truth was provided by high-precision GPS. In these complex conditions, traditional methods exhibited wild fluctuations in altitude estimation once the target dropped below a 2.5° elevation. In contrast, the proposed algorithm maintained a much tighter error margin. Specifically, over 56% of its estimates fell within a 260m error window, a significant improvement over the 35-49% achieved by competing methods.
Conclusion
The proposed sparse representation approach provides a robust and accurate solution for low-angle radar tracking in complex environments. By utilizing physics-informed dictionary atoms and an adaptive refinement strategy, it overcomes the limitations of traditional flat-earth models and subspace methods. While the computational complexity is higher than simple MUSIC-based approaches, the gains in accuracy and reliability—especially in limited snapshot and complex terrain scenarios—make it a powerful tool for modern radar systems.
Summary updated