An Adaptive Beam-Steering dToF LiDAR System Using Addressable Multi-Channel VCSEL Transmitter, 128 × 80 SPAD Sensor, and ML-Based Edge-Computing Object Detection
- 4 April 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 72 (5) , 2089-2102
Abstract
In this work, a solid-state direct time-of-flight (dToF) and adaptive beam-steering Light Detection and Ranging (LiDAR) system is proposed for machine learning (ML) based object detection. To leverage the capabilities of software and hardware, a co-optimization design from a neural network based algorithm to the architecture of transmitter, receiver and optical components is realized. Firstly, an object detection neural network is proposed for the depth-only input algorithm, which indicates the Region of Interest (ROI) in the illuminating field and gives hints of opened scan channels in the next two frames to decrease the total cost of the laser driver and sensor array. Next, the proposed network utilizes the Cross-Stage-Patrial (CSP) block to replace the residual structure in the backbone to achieve a lightweight performance and is implemented on the NVIDIA-Jetson to verify the system-level adaptive beam steering feature. To realize the smart working mode, a customized multi-channel and addressable TX is designed for adaptive and optical control to save power consumption and extend the ranging distance. At the same time, a $128\times 80$ resolution RX which consists of Single-Photon Avalanche Diodes (SPADs) and column-wise Time-to-Digital Converter (TDC) is incorporated to capture the returned photons for combining sub-regions into an entire depth map. Next, to customize the specific scanning mechanism, for the optical setup, a cylindrical lens array is designed to reshape the laser beam, which matches the pattern of the transmitter to illuminate different targeted objects. Both the laser driver chip and the sensor chip with a $128\times 80$ SPAD array are fabricated in the 180-nm Bipolar-CMOS-DMOS (BCD) process. Finally, the laser driver chip realizes the power of 5 W with an adjustable pulse width of 1.5 ns and the SPAD array integrates the depth accuracy of 5 cm at 15 m. Due to that the neural network realizes an accuracy up to 0.8, a low-power solid-state LiDAR prototype with adaptive beam steering is demonstrated.Keywords
Funding Information
- Fundamental Research Funds for the Central Universities (TJ)
This publication has 41 references indexed in Scilit:
- A Sub-ns Pulsed VCSEL Driver with Real-Time Adaptive Current Control, Integrated Boost Switching Regulator and Class-1 Laser Eye Safety2023
- dToF LIDAR System Using Addressable Multi-Channel VCSEL Transmitter, 128x80 SPAD Sensor, and ML-Based Object Detection for Adaptive Beam-Steering2023
- Towards a solid‐state light detection and ranging system using holographic illumination and time‐of‐flight image sensingJournal of the Society for Information Display, 2022
- RepVGG: Making VGG-style ConvNets Great Again2021
- nuScenes: A Multimodal Dataset for Autonomous Driving2020
- Deep learning optimized single-pixel LiDARApplied Physics Letters, 2019
- PointPillars: Fast Encoders for Object Detection From Point Clouds2019
- Object Classification using Deep Learning on Extremely Low-Resolution Time-of-Flight Data2018
- Neural network identification of people hidden from view with a single-pixel, single-photon detectorScientific Reports, 2018
- GaN laser driver switching 30 A in the sub-nanosecond range2016