A 128×96 Multimodal Flash LiDAR SPAD Imager with Object Segmentation Latency of 18μs Based on Compute-Near-Sensor Ising Annealing Machine

📋 论文概要

本文提出了一款128×96多模态闪光LiDAR SPAD图像传感器，集成了近传感器计算伊辛模型退火处理器，用于动态目标分割。通过将分割问题转化为组合优化问题，并利用多模态SPAD信息加速伊辛自旋迭代，实现了18微秒的分割延迟。

💡 主要创新点

• 首次将伊辛模型退火处理器集成到SPAD LiDAR图像传感器中，实现近传感器计算，大幅降低目标分割延迟。
• 提出高带宽输出同时写入方案，支持哈密顿系数的实时更新，适应动态场景。
• 利用多模态SPAD信息（如强度、深度）加速伊辛自旋迭代，提升分割效率。

🏷 关键词

📄 原文摘要

*Equally-Credited Authors (ECAs) Abstract This paper presents a 128×96 multimodal flash LiDAR SPAD imager integrated with a compute-near-sensor Ising-model annealing processor for dynamic object segmentation, framed as a combinatorial optimization problem (COP). The system enables real-time updates of Hamiltonian coefficients using a high-bandwidth output-while-write scheme, and accelerates Ising spin iterations by exploiting multimodal SPAD information, achieving a segmentation latency of 18µs. Single-photon avalanche diode (SPAD) imagers offer single-photon sensitivity, picosecondlevel timing resolution, and intrinsic time-of-flight depth sensing capabilities, enabling robust 2D/3D perception in low-light and high-dynamic-range scenarios. These attributes render SPADs highly suitable for real-time sensing applications including autonomous driving, 3D mapping, and augmented reality (AR) [1-3]. In many scenarios, object

👥 作者与机构

Jingyi Wang*, Tao Hong*, Bu Chen*, Zhangcheng Huang, Qi Liu, Ming Liu

Fudan University, Shanghai, China