Deep Neural Network-Based Phase-Modulated Continuous-Wave LiDAR.

Area of Science:

• Optics and Photonics
• Computer Vision and Machine Learning
• Robotics and Autonomous Systems

Background:

• Light Detection and Ranging (LiDAR) offers high accuracy and resolution, crucial for various applications.
• Phase-modulated continuous-wave (PhMCW) LiDAR provides advantages like low power consumption and high precision.
• Degradation in signal-to-noise ratio (SNR) severely impacts traditional time-of-flight extraction methods in LiDAR.

Purpose of the Study:

• To introduce a novel deep neural network (DNN) approach for robust time-of-flight measurement in PhMCW LiDAR.
• To evaluate the DNN method's performance under varying distance resolutions and low SNR conditions.
• To demonstrate the DNN's capability in reconstructing detailed 3D point clouds from simulated complex scenes.

Main Methods:

• Development of a deep neural network model designed to directly process LiDAR signals.
• Simulation of a 6-meter range scene including diverse objects (vehicle, tree, house) and background.
• Analysis of recognition accuracy and performance metrics at a 0.1 m distance resolution and SNR as low as 2.

Main Results:

• The proposed DNN method achieved an 81.4% recognition accuracy under challenging low SNR (2) and 0.1 m resolution.
• Simulated point cloud reconstruction exhibited high fidelity with clear object contours and restored features.
• Precise distance measurements (4.73 cm, 6.00 cm, 7.19 cm) demonstrated the method's excellent performance.

Conclusions:

• Deep neural networks offer a viable and effective solution for processing LiDAR signals and extracting time-of-flight.
• This work represents the first application of neural networks for direct LiDAR signal processing and time-of-flight extraction.
• The DNN method shows significant potential for improving LiDAR performance in low SNR environments and complex scenarios.