Neuromorphic computing for compact LiDAR systems

Dennis Delic, Saeed Afshar

Research output: Chapter in Book / Conference PaperChapter

Abstract

Driven by both defense and consumer markets, the impact of More-than-Moore philosophy is seeing substantial progress in Light Detection and Ranging (LiDAR)-based technologies for remote sensing applications. Single-photon LiDAR (SPL) is a promising class of LiDAR technology that relies on single-photon detectors (SPDs) that can achieve extremely sensitive detection levels and performance, providing new sensing capabilities for object detection and identification. Single-photon avalanche diodes (SPADs) are low-cost, scalable, solid-state SPDs that are compatible with existing complementary metal-oxide-semiconductor (CMOS) integrated circuit (IC) design techniques; this makes them an ideal candidate for imaging-based SPL systems. Various SPAD-based SPL configurations are reviewed; a specific type known as 3D flash LiDAR represents a compact system that can acquire a 3D depth-resolved image of a scene instantaneously, can measure fine spatial details of objects in the scene, and is capable of imaging through partial obscurants. These abilities allow high-speed image capture and classification of fast-moving objects, especially those that are hard to see and concealed in highly cluttered backgrounds, capabilities that are highly desirable for military applications. The key to improving the performance and functionality of 3D flash LiDAR systems for long-range sensing and tracking applications is the SPAD array also known as a focal-plane array (FPA) sensor. The FPA includes both an array of SPAD detectors and CMOS-compatible electronics and/or readout integrated circuits (ROICs) which collectively can be considered a smart-pixel array. The number of smart pixels on the FPA determines both the field of view (FOV) for the 3D flash LiDAR system and the amount of scene data generated. This data needs to be processed at a fast-enough rate to ensure minimal latency times for real-time operation; this is especially important when imaging and tracking moving objects. To address the FOV limitations, SPAD FPA imaging sensors packing more pixels per unit size is achievable by using advanced 3D stacking assembly and integration technologies, while new ROIC capabilities employing machine vision computation methods and event-based neuromorphic processing solutions can overcome the data processing bottleneck problem; all these techniques are discussed. Novel neuromorphic event-based SPAD FPA sensor architectures are presented, showing that event-based methods are superior to the frame-based approaches in terms of reducing the output data rate, improving imaging responsivity, being less susceptible to noise, and enabling fast processing speed for real-time detection and classification. Future SPAD FPA image sensors that can integrate neuromorphic computing methods with artificially intelligent (AI) techniques will reduce power consumption and data flow and offer new detection and learning processing capabilities. Advanced 3D manufacturing and integration techniques will see further increase in performance, beyond what is possible with Moore’s law. The next generation of smart SPL imaging sensors will have the potential to defeat stealth or low observable (LO) technologies; improve surveillance imaging from space, in air, and underwater; sense fast-moving objects partially concealed behind obscurants with low detectability; and identify objects in very low light conditions and/or hidden in background clutter.

Original languageEnglish
Title of host publicationMore-than-Moore Devices and Integration for Semiconductors
EditorsFrancesca Iacopi, Francis Balestra
Place of PublicationSwitzerland
PublisherSpringer
Pages191-240
Number of pages50
ISBN (Electronic)9783031216107
ISBN (Print)9783031216091
DOIs
Publication statusPublished - 1 Jan 2023

Fingerprint

Dive into the research topics of 'Neuromorphic computing for compact LiDAR systems'. Together they form a unique fingerprint.

Cite this