Abstract
A novel framework of signal extraction is proposed based on improved singular spectrum analysis (SSA) and compressed sensing (CS) for laser radar. The improved signal selection method, which locates a cut-off point on the singular spectrum curve, is fused into the SSA to reduce the noise. Afterward, the CS in the wavelet domain is employed to acquire the observation vector, and the improved reconstruction algorithm with a hard threshold denoising method reduces the noise furthermore. The simulations and experiments are performed to verify the superiority of the proposed algorithm compared with other algorithms. The proposed algorithm can enhance a signal-to-noise ratio while reducing the distortion of the received signal.
The Challenge of Low SNR in LiDAR
Direct detection laser radar (LiDAR) systems calculate distance by measuring the time-of-flight of a laser pulse. The precision of these measurements depends entirely on the system's ability to accurately identify the peak of the returning echo signal. However, in real-world applications, these signals are frequently buried in noise, leading to a low Signal-to-Noise Ratio (SNR). Traditional analog detection methods, such as peak or leading-edge detection, often fail under these conditions.
While digital denoising techniques like wavelet transforms and Singular Spectrum Analysis (SSA) offer solutions, they come with trade-offs. Wavelet soft-thresholding tends to smooth out noise but also reduces the peak amplitude of the signal, while hard-thresholding can introduce artificial oscillations. Standard SSA struggles to precisely define the boundary between signal and noise singular values. Furthermore, Compressed Sensing (CS) typically requires prior knowledge of signal sparsity, which is often unknown for natural LiDAR echoes.
Proposed Methodology
To address these limitations, the researchers developed a two-stage digital signal processing framework that combines Improved SSA (ISSA) with CS in the wavelet domain.
Stage 1: Improved Singular Spectrum Analysis (ISSA)
The first stage involves pre-denoising the signal using ISSA. Unlike standard SSA, which may rely on simple peak-picking in the curvature spectrum, the proposed ISSA uses a novel cut-off selection method based on energy ratios and curvature thresholds. This approach more robustly separates the true echo signal from the background noise floor.
Stage 2: Compressed Sensing in the Wavelet Domain
The pre-denoised signal is then transformed into the wavelet domain for further refinement. The researchers utilized an Improved Generalized Orthogonal Matching Pursuit (IGOMP) algorithm for reconstruction. Key innovations in this stage include:
- Adaptive Sparsity Determination: A discrete curvature analysis method is used to determine the $K$-sparsity of the signal on the fly, solving the "unknown sparsity" problem.
- Post-Reconstruction Cleanup: To eliminate residual artifacts or "wrong atoms" selected by the greedy CS algorithm, a hard threshold is applied to the wavelet detail coefficients after reconstruction.
Experimental Setup
The framework was validated using both synthetic and real-world data. Synthetic signals were modeled as Gaussian functions with varying levels of added noise. Real-world data was collected using a custom optical setup.
The experimental system featured a 1064nm pulse laser, an avalanche photodiode (APD) receiver, and a high-speed oscilloscope sampling at 1GHz. Targets were positioned at a distance of 5 meters to capture authentic echo waveforms corrupted by electromagnetic interference and power supply noise.
Key Results and Insights
The proposed ISSA+CS method demonstrated significant advantages over traditional denoising techniques:
- Superior SNR Enhancement: In simulations with 10mV RMS noise, the proposed method achieved an SNR of 25.84 dB, outperforming standard hard thresholding (19.87 dB), soft thresholding (20.74 dB), and standard CS methods (22.44 dB).
- Timing Precision: The method maintained a Peak Time Error (PTE) of 0 ns across all tested noise levels. In contrast, competing methods showed drifts of 1 to 8 ns, which would translate to significant physical ranging errors.
- Amplitude Preservation: Unlike soft-thresholding methods that flatten signal peaks, the ISSA+CS approach preserves the exact peak voltage and timing, which is critical for accurate LiDAR ranging.
- Real-World Robustness: Experimental results confirmed the algorithm's ability to remove coupled impulse noise from power supplies while maintaining the integrity of the signal's shape.
Conclusion and Future Work
The proposed two-stage framework provides a robust solution for extracting weak LiDAR echo signals in high-noise environments. By combining the strengths of ISSA and CS with adaptive sparsity determination, the system achieves high SNR and timing accuracy without signal distortion. Future research will focus on extending this framework to multi-target scenarios, where overlapping waveforms and varying reflection intensities present even greater challenges for signal extraction.
Summary updated