First-Arrival Differential Counting for SPAD Array Design

Mel J. White, Tianyi Zhang, Akshat Dave, Shahaboddin Ghajari, Alyosha C. Molnar, Ashok Veeraraghavan2023 · DOI

TL;DR

This paper introduces a novel "First-Arrival Differential Counting" (FAD) architecture for Single-Photon Avalanche Diode (SPAD) arrays. This design significantly reduces the need for complex per-pixel timing circuitry, enabling higher pixel density and lower data throughput, while achieving high dynamic range imaging and high-resolution LiDAR.

What Problem Does This Paper Address?

Traditional SPAD arrays, while excellent for low-light conditions, face challenges when dealing with high photon flux or when high spatial resolution is required for applications like 3D imaging (LiDAR). To capture precise timing information for LiDAR, each SPAD pixel typically needs a Time-to-Digital Converter (TDC). These TDCs are large, reducing the "fill factor" (the proportion of the pixel area dedicated to sensing) and increasing data bandwidth, which limits resolution and frame rates. For high dynamic range (HDR) imaging, SPADs can saturate quickly, losing information. This paper tackles these limitations by proposing a new readout architecture that captures relative intensity or timing information between pixels, rather than absolute values, thus simplifying the circuitry and improving performance.

Methods

The core innovation is the "FAD unit," which uses a simple SR (set-reset) latch. This latch captures the first photon arrival between two SPAD pixels within a given time window. Instead of recording the exact arrival time, it records which of the two SPADs received the first photon. This information is then fed into an up/down counter.

  • Differential Counting: By comparing the "first arrival" events between pairs of pixels, the system can infer relative intensity differences (for HDR imaging) or relative time-of-flight differences (for LiDAR).
  • Reduced Circuitry: The FAD unit is significantly smaller than a TDC, allowing for a much higher fill factor and denser arrays. Crucially, the counters and other digital logic can be placed off-pixel, further reducing complexity.
  • Architecture: The paper explores two main connectivity schemes:
  • Haar Grouping: Hierarchical grouping of pixels, suitable for HDR imaging, which compresses information efficiently.
  • Nearest Neighbor: Connecting adjacent pixels, ideal for capturing local gradients and surface normals in LiDAR applications.
  • Dual Photon Event Handling: An optional AND gate can be used to filter for "dual arrival" events (where both SPADs in a pair detect a photon), which is crucial for accurate LiDAR depth estimation by rejecting background noise.

Datasets

The paper uses simulated data for its demonstrations. For HDR imaging, it references an original HDR image from [47]. For LiDAR, it uses simulated scenes and ground truth depth maps for evaluation. No new datasets were introduced.

Key Insights and Ideas

  • First Arrival is Rich Information: The paper's key insight is that for many applications, the first photon arrival within a time window contains sufficient information, eliminating the need for precise timestamps.
  • Differential Sensing for HDR: By comparing photon arrivals between pixels, the FAD architecture can preserve information even when individual pixel counters would saturate, enabling high dynamic range imaging with a single exposure.
  • FAD-LiDAR: The differential timing information from FAD units, combined with a few sparse TDCs for absolute depth anchoring, allows for high-resolution, single-shot 3D imaging (Flash LiDAR) with significantly reduced data throughput compared to traditional per-pixel TDC approaches.
  • Scalability: The simplified FAD unit architecture is highly scalable, paving the way for much larger and denser SPAD arrays.

Key Results

  • HDR Imaging: Simulations show that the FAD-SPAD approach preserves gradient information and dynamic range even when individual pixel counters saturate, outperforming traditional methods that lose contrast.
  • LiDAR: The FAD-LiDAR system demonstrates high-quality 3D reconstruction and surface normal estimation, comparable to scanning LiDAR, with improved performance over conventional flash LiDAR baselines at matched data bandwidth.
  • Fill Factor: The prototype achieved a 36% fill factor, significantly higher than typical per-pixel TDC SPAD arrays.
  • Data Throughput: The differential nature of the FAD approach inherently reduces data throughput, enabling higher resolution and frame rates.

Limitations and Future Work

  • TDC Sparsity for LiDAR: While FAD reduces the need for TDCs, some sparse TDCs are still required for absolute depth anchoring in LiDAR applications. The paper explores the trade-off between TDC density and depth accuracy.
  • Albedo Variation: In LiDAR, significant albedo variation between pixels can introduce bias and increase depth estimation errors, particularly when the albedo ratio exceeds 0.5.
  • Jitter: SPAD jitter affects depth resolution, with larger jitter enabling wider gradient detection but reducing depth accuracy.
  • Future Work: The authors suggest exploring other differential connectivity schemes for compressed sensing, adaptive sensing, and developing larger, denser SPAD arrays with high photon detection probability. They also highlight the potential for FAD's differential sensing to enhance contrast in bioimaging applications.

Summary updated

BibTeX
@misc{white2023first,
  author = {Mel J. White and Tianyi Zhang and Akshat Dave and Shahaboddin Ghajari and Alyosha C. Molnar and Ashok Veeraraghavan},
  title = {{First-Arrival Differential Counting for SPAD Array Design}},
  year = {2023},
  doi = {10.3390/s23239445},
}
Figure