TL;DR
This paper introduces a novel deep neural network (DNN) approach for directly processing Phase-Modulated Continuous-Wave (PhMCW) LiDAR signals to measure time-of-flight (ToF). By training a ResNet152 model to classify pulse widths, the researchers achieve accurate distance measurements even at low signal-to-noise ratios (SNRs) and high resolutions, overcoming limitations of traditional methods.
What Problem Does This Paper Address?
Traditional LiDAR systems, particularly Time-of-Flight (ToF) and Frequency-Modulated Continuous-Wave (FMCW) schemes, face challenges. ToF LiDARs can have eye-safety issues and lack velocity measurement capabilities, while FMCW LiDARs struggle with laser linearity and environmental sensitivity. PhMCW LiDAR offers advantages like low power and high precision but suffers from a decreasing signal-to-noise ratio (SNR) at longer distances or with low-reflectivity targets. This low SNR severely degrades the signal waveform, making traditional methods for extracting the time-of-flight (by measuring pulse width) unreliable. The paper aims to address this by developing a more robust method for pulse width measurement in PhMCW LiDAR, thereby enhancing its long-distance ranging capabilities.
Methods
The core of this work is the application of a deep neural network, specifically a ResNet152 architecture, for pulse width recognition in PhMCW LiDAR signals. Instead of traditional methods that rely on detecting pulse edges, the DNN directly analyzes the digitized intermediate frequency (IF) signal waveform. The approach taken is classification-based, where pulse widths are categorized into discrete values corresponding to specific distances. This is preferred over regression because it provides stable measurements within a defined resolution, which is more practical for LiDAR applications. The ResNet architecture was chosen for its ability to maintain high accuracy with increased depth, which is crucial for handling noisy signals and distinguishing between very similar waveforms at high resolutions. The network is trained to map time-domain waveforms to their corresponding pulse widths, which are then converted to distances using the formula $d = (c \times t)/2$.
Datasets
The study utilizes simulated IF waveforms for training and testing. These simulated datasets are generated with varying pulse widths, distance resolutions, and signal-to-noise ratios (SNRs). The paper does not mention the use of publicly available real-world LiDAR datasets. For the point cloud imaging demonstration, a simulated scene was created, containing a vehicle, a tree, a house, and a background at different distances, with simulated waveforms generated based on these distances and calculated SNRs.
Key Insights and Ideas
- DNNs for Direct Signal Processing: This paper pioneers the use of neural networks to directly process raw LiDAR IF signals for distance extraction, moving beyond traditional signal processing techniques.
- Robustness to Low SNR: The DNN approach demonstrates remarkable resilience to low SNRs, achieving acceptable accuracy even when traditional methods would fail. This is a significant advancement for extending LiDAR's operational range.
- Classification for Practical Resolution: Employing a classification strategy for pulse width measurement is shown to be more practical for LiDAR applications than regression, as it yields results within a defined, achievable resolution.
- ResNet's Suitability for LiDAR: The ResNet152 architecture proves effective in handling the challenges of noisy and similar waveforms, demonstrating the power of deep residual learning in this domain.
Key Results
- High Accuracy at High Resolution: The network achieves an accuracy of 90.2% at a high resolution of 0.1 m.
- Performance at Low SNR: Even with an SNR as low as 2, the network maintains an accuracy of 81.4% at a 0.1 m resolution.
- Excellent Point Cloud Reconstruction: A simulated point cloud imaging demonstration showed that the reconstructed scene had good fidelity, clear object contours, and restored features, with object precisions comparable to commercial LiDAR products (e.g., 4.73 cm for a vehicle, 6.00 cm for a tree/house, and 7.19 cm for the background).
- Measurement Speed: The trained network can measure distance in approximately 15 ms per point, meeting real-time measurement requirements.
Limitations and Future Work
- Simulation-Based Training: The current work relies on simulated data for training. While the simulation is designed to be representative, real-world data may introduce unforeseen complexities. The authors suggest that future work could involve training with real data.
- Limited Range in Demonstration: The point cloud imaging demonstration used a limited range (up to 6 m) and a specific set of SNRs (1.64 to 5.00). While the network was trained for an SNR of 2, the actual SNRs in the demonstration were slightly higher.
- Computational Cost: Training the network currently takes approximately 10 days on a CPU. The authors note that GPUs, FPGAs, or photonic AI chips could significantly accelerate this process.
Future work could focus on extending the operational range, incorporating real-world data for training, and exploring hardware acceleration for faster deployment.
Summary updated