Time-Resolved Photon Counting Cameras

• Time-resolved single-photon counting cameras are imaging systems that convert individual photon arrivals into spatially indexed temporal observables for precise measurements.
• They integrate sensor architectures like SPAD arrays, TCSPC modules, and intensified cameras to optimize timing resolution and photon detection efficiency.
• Applications include fluorescence lifetime imaging, photon-counting depth imaging, and quantum-optical correlation measurements through advanced calibration and reconstruction techniques.

Time-resolved single-photon counting cameras are imaging systems that register individual photon events and preserve temporal information at the level of per-pixel timestamps, gated histograms, or equivalent time-binned measurements. In the literature, this category spans CMOS SPAD arrays with on-chip gating or timestamping, intensified event-driven cameras based on Timepix readout, single-pixel and line-scanning TCSPC instruments, and non-standard architectures that translate inaccessible spectral bands into silicon-compatible detection or exploit on-chip nonlinear absorption for optical gating (Bruschini et al., 2019, Nomerotski et al., 2022, Fang et al., 1 Jun 2026, Dai et al., 2017). Across these implementations, the defining operation is the conversion of sparse photon arrivals into spatially indexed temporal observables—arrival times, histograms, gate counts, inter-photon delays, or time-of-flight distributions—that support fluorescence lifetime estimation, photon-counting depth imaging, diffuse optical localization, ultrafast thermal imaging, and quantum-optical correlation measurements (Wu et al., 2024, Ingle et al., 2021, Abelson et al., 5 May 2026).

1. Detection principles and temporal observables

The canonical detector element in many time-resolved single-photon cameras is the SPAD operated in Geiger mode, where a single photo-generated carrier can trigger a self-sustaining avalanche and produce a digital pulse. Array implementations combine this binary event generation with quenching, gating, counters, or TDCs, so that a photon is represented not by an analog charge packet but by a digital event tied to a time reference (Bruschini et al., 2019). In TCSPC use, the measured histogram is commonly modeled as an instrument-response-limited scene response plus background. For diffuse-light imaging, the reported model is

$M(t) = (IRF * S)(t) + B,$

where $S(t)$ captures the scene response and $B$ accounts for ambient and dark counts (McShane et al., 2022). In fluorescence applications, the corresponding form

$S(t) = (IRF * I)(t) + B$

is paired with exponential decay models and iterative reconvolution fitting (Wahl et al., 2020).

Time-of-flight imaging adopts the same temporal primitives but interprets delay as range. The basic conversion is

$d = \frac{c\,t}{2},$

with the same relation reused in direct ToF, photon-counting 3D imaging, and NLOS backprojection (Dai et al., 2017, Wu et al., 2024). The significance of timing performance follows immediately: a 5 ps instrument response in an optically gated silicon EMCCD corresponds to $\Delta z \approx 0.75$ mm in ToF depth precision (Fang et al., 1 Jun 2026), whereas a miniaturized TCSPC module with 27.4 ps minimum root-mean-square time resolution does not itself guarantee millimetric NLOS resolution because the overall depth resolution is limited by the full IRF and geometry (Wu et al., 2024).

Not all time-resolved single-photon cameras detect photons by direct avalanche sensing. A telecom-band silicon EMCCD can instead exploit non-degenerate two-photon absorption, with the condition

$E_s + E_p \ge E_g,$

so that a 1550 nm signal photon and a 3070 nm pump photon jointly generate carriers in silicon; choosing the pump such that $2E_p < E_g$ suppresses pump-only degenerate two-photon absorption and harmonic background (Fang et al., 1 Jun 2026). Mid-infrared imaging can also be transferred into the silicon-SPAD domain by broadband adiabatic frequency upconversion, obeying

$$\frac{1}{\lambda_{SFG}} = \frac{1}{\lambda_{pump}} + \frac{1}{\lambda_{MIR}},$$

thereby preserving the SPAD camera’s photon-counting and gated timing functionality while moving sensitivity into the 2–5 µm band (Abelson et al., 5 May 2026). A different extension replaces active illumination altogether: passive inter-photon imaging uses the inter-photon delay between successive detections as an intensity cue, with the ideal inter-arrival law

$p(t \mid \lambda) = \lambda e^{-\lambda t},$

and a dead-time-aware flux estimator derived from the first and last timestamps and the photon count (Ingle et al., 2021).

2. Sensor architectures and readout organizations

Architecturally, time-resolved single-photon cameras range from dense 2D SPAD matrices to scanning systems and hybrid detector chains. The SPAD-array review emphasizes a design spectrum that includes basic SPAD-plus-discriminator pixels, in-pixel counters, in-pixel TACs or TDCs, shared column TDCs, multi-gate counters, masking, and sub-nanosecond gating, with 1D arrays often moving electronics off-pixel to recover fill factor and 2D arrays trading timing resources against pitch and data rate (Bruschini et al., 2019). Representative sensors in that survey include SwissSPAD, SwissSPAD2, MEGAFRAME32, MEGAFRAME128, Piccolo, and several Raman and PET-oriented architectures with timing bin sizes in the 40–100 ps regime (Bruschini et al., 2019).

A distinct line-scanning realization uses a 512×1 CMOS SPAD line array, each pixel aggregating 16 SPADs in an 8×2 sub-array, with a motorized mirror sweeping a second field dimension. In that configuration, 512 sequential slices reconstruct a 512×512 image, or approximately 0.26 Mpixel, over a 30 cm by 30 cm field of view at 61 cm stand-off distance (McShane et al., 2022). The same work emphasizes that the large physical extent of the line sensor enables larger optics and higher etendue than compact 2D SPAD arrays (McShane et al., 2022).

Intensified event-driven cameras form another major branch. The intensified Tpx3Cam couples a 256×256 Timepix3 sensor at 55×55 µm² pitch to a microchannel-plate image intensifier and a fast P47 phosphor. Each hit pixel records Time-of-Arrival and Time-over-Threshold, sparse readout is sustained at 80 Mpixel/s, and single-photon events are reconstructed as multi-pixel clusters using recursive grouping within a 300 ns window (Nomerotski et al., 2022). The Timepix4-based successor expands the matrix to 512×448 pixels at the same 55 µm pitch, reduces the ToA bin to 195 ps, and preserves data-driven multi-hit acquisition while allowing direct optical detection or intensified single-photon operation (Hogenbirk et al., 18 Sep 2025).

Silicon EMCCD architectures can also be time-resolved when the temporal gate is optical rather than electronic. In the NDTPA implementation, a commercial Andor iXon Ultra 888 EMCCD acts as both imager and nonlinear absorber: pump and signal pulses coincide in the silicon depletion region, generate carriers through on-chip nonlinearity, and the charges are integrated, multiplied by an EM gain of approximately 1000, and read out at 30 MHz (Fang et al., 1 Jun 2026). Because the gate is set by the optical pulse overlap rather than a conventional shutter, the temporal resolution is set by the pump–signal cross-correlation rather than the tens-of-nanoseconds scale typical of electronic gating (Fang et al., 1 Jun 2026).

Scanning and single-pixel systems remain important where array formats are limited or a different trade-off is desired. AC3DI uses a DMD, structured illumination, a single-pixel photon-counting detector, and a PicoHarp 300 to recover intensity and depth from two linear measurements per projected Hadamard pattern: total photon count and sum of photon TOFs (Dai et al., 2017). Time-multiplexed near-infrared detectors implement a related principle in the temporal domain by splitting one spatial channel into 16 or 32 delayed bins interrogated by gated InGaAs/InP APDs, effectively producing a 1×N time-resolved single-photon camera (Eraerds et al., 2010). At the architectural frontier, a 110 nm CIS SPAD demonstrator embeds FPGA-like LUT logic under reconfigurable macropixels so that timestamping and photon counting can be mixed with coincidence logic, weighted sums, and ROI-dependent mode selection directly on-chip (Milanese et al., 20 Nov 2025).

3. Timing performance, calibration, and statistical reconstruction

Temporal performance in this field is a compound property of detector jitter, gate width, TDC quantization, optics, and calibration. The 512×1 line-scan SPAD camera operates with approximately 50 ps time stamping and approximately 150 ps per-pixel jitter; the effective IRF is limited by the sensor jitter and the 96 ps laser pulse width, giving a combined response on the order of a few hundred picoseconds (McShane et al., 2022). The miniaturized four-channel TCSPC module reports a 10 ps bin size, 27.4 ps minimum root-mean-square time resolution, 64.5 ps minimum FWHM time resolution, max DNL 0.17 LSB, and max INL 8.49 LSB, but in NLOS imaging the system resolution is dominated by the 110 ps FWHM detector jitter and geometric broadening rather than by TDC quantization (Wu et al., 2024).

Event-driven intensified cameras require explicit time-walk compensation because threshold crossing depends on pulse amplitude. In the Tpx3Cam case, the Timepix3 discriminator exhibits time-walk up to approximately 100 ns for signals near threshold, so a camera-specific ToT correction function is applied; after correction, simultaneous SPDC photons yield an inter-photon timing RMS of approximately 3.3 ns for the pair, or approximately 2.4 ns per photon under equal-contribution assumptions (Nomerotski et al., 2022). Timepix4 retains the same ToA/ToT paradigm but benefits from a 195 ps ToA bin; measured RMS timing reaches 0.318 ns without intensifier and 0.272 ns with a tight ToT slice in direct detection, while intensified single-photon operation yields 0.83 ns, 0.55 ns, or 1.42 ns depending on amplitude selection and time-walk correction (Hogenbirk et al., 18 Sep 2025).

Optically gated silicon NDTPA cameras sit in a different regime. There, the per-pixel cross-correlation of counts versus pump–signal delay gives an instrument response width of approximately 5 ps, with electronic readout jitter and carrier transport negligible compared to the optical gate (Fang et al., 1 Jun 2026). At still shorter effective resolutions, time-magnified photon counting inserts a temporal imaging system before the detector so that the measured timing uncertainty is reduced after demagnification. With a measured magnification of approximately 130.5, a standard near-infrared SPAD and commercial TDC achieve an effective Gaussian-equivalent FWHM of 550 fs, resolve ultrashort pulses with a 130-fs pulsewidth difference at a 22-fs accuracy, and suppress range walk error by 99.2% in photon-counting ToF imaging (Li et al., 2021).

Calibration strategies follow the architecture. TCSPC and FLIM systems measure the IRF and incorporate it through iterative reconvolution or other forward models (Wahl et al., 2020). Line-scanning diffuse imagers measure the IRF by direct illumination and then rely on early-photon gating rather than numerical deconvolution (McShane et al., 2022). Timepix systems calibrate ToT-to-time correction, centroid clusters with ToT weights, and in some cases apply per-pixel time-offset calibration (Nomerotski et al., 2022, Hogenbirk et al., 18 Sep 2025). The multi-channel MultiHarp-class timing electronics extend these ideas to many detectors at once by providing 80 ps base resolution, 650 ps per-channel dead time, T2 and T3 event formats, hardware sorting and histogramming, and optional White Rabbit synchronization with 89 ps rms jitter measured between two taggers over 5 km fiber (Wahl et al., 2020).

4. Spectral coverage and modality extensions

Although silicon SPAD arrays matured first in the visible and near-infrared, time-resolved single-photon counting cameras now cover a broader spectral range through detector conversion, nonlinear absorption, and alternative detector technologies. The review literature frames visible and NIR SPAD cameras as the dominant CMOS platform, with applications in FLIM, Raman, NIROT, PET, and correlation imaging (Bruschini et al., 2019). In the telecom band, silicon can be repurposed by NDTPA rather than replaced: using a 3070 nm pulsed mid-infrared pump together with a 1550 nm signal, a silicon EMCCD reaches single-photon sensitivity around 1 photon/pixel/pulse, greater than 30-fold enhancement of photon-counting rate over degenerate two-photon absorption, approximately 5 ps temporal resolution, and approximately 13 µm spatial resolution without phase matching (Fang et al., 1 Jun 2026).

Mid-infrared operation has been extended by upconversion rather than by direct mid-IR SPAD fabrication. A broadband adiabatic frequency-upconversion stage in APLN maps 2–5 µm photons into 680–880 nm and feeds a 512×512 silicon SPAD array with integrated global gating. That system demonstrates spectrally resolved mid-IR imaging, nanosecond-scale thermal dynamics via weak mid-IR blackbody emission, full-field imaging at up to 60,000 frames per second, and 1-bit sensor operation up to 100,000 fps under suitable flux (Abelson et al., 5 May 2026). The timing in that architecture is gate-defined rather than TDC-defined, so the effective temporal slices are nanosecond-scale and were sampled in 5 ns delay increments (Abelson et al., 5 May 2026).

Other spectral extensions rely on non-silicon detectors. A free-space-coupled superconducting microstrip single-photon detector provides a 260 µm-diameter active area, about 171 ps FWHM system jitter at 1550 nm, and low-dark-count operation at approximately 2.0 K, enabling photon-counting ToF scanning with a large-area detector that relaxes alignment relative to fiber-coupled SNSPDs (Wang et al., 2024). A compact NLOS system instead pairs a 1550 nm InGaAs/InP single-photon detector, operated in 10 ns gated mode at 10 MHz, with a miniaturized TCSPC module and galvanometer indexing to reconstruct hidden objects at 5 m range (Wu et al., 2024). Time-multiplexed near-infrared APD systems extend 1100–1650 nm counting into 16- or 32-bin temporal vectors and report single-shot energy resolution on the attojoule level at 1550 nm (Eraerds et al., 2010).

This range of examples suggests that “time-resolved single-photon counting camera” is better understood as a measurement class than as a single detector family. Direct avalanche arrays, intensified optical relays, superconducting detectors, gated InGaAs modules, upconversion front-ends, and optically gated silicon imagers all satisfy the same operational criterion when they return spatially or temporally indexed photon events with preserved arrival-time information.

5. Imaging workflows and application domains

Application workflows differ widely, but most reduce to acquiring sparse temporal observables and reconstructing structure, decay, or correlation from them. In diffuse-light localization, the line-scanning SPAD system determines the TCSPC histogram peak for each emission point, defines “early time” when the signal first reaches 10% of the peak, extracts the image at that early time bin, and computes the centroid of the bright region to estimate source location. In a milk phantom, this early-photon selection localizes 18 discrete emission points with average deviation approximately 0.51 cm, whereas summed all-photon images overlay poorly with the true path (McShane et al., 2022).

Photon-counting 3D imaging with a single-pixel detector adopts a different reconstruction. AC3DI records, for each Hadamard pattern, both photon count and TOF sum, performs inverse Hadamard demultiplexing to recover intensity and a “modulated image,” and then estimates depth by ratio. Wavelet-tree prediction marks edge regions for refinement at progressively finer scales, so that a 512×512 3D image can be acquired with practical time as low as 17 seconds at 5% of the full measurement set and reconstructed in under 1 second (Dai et al., 2017). NLOS imaging with the compact TCSPC module uses the confocal ellipse constraint and standard elliptical backprojection, reaching 6.3 cm lateral resolution and 2.3 cm depth resolution at 5 m range with 1 ms pixel dwell time (Wu et al., 2024).

The same timing machinery supports FLIM and lifetime-sensitive microscopy. The EMCCD NDTPA camera is explicitly positioned for low-light fluorescence lifetime microscopy because coincident optical gating yields approximately 5 ps temporal resolution without electronic jitter limitations (Fang et al., 1 Jun 2026). Multi-channel TCSPC electronics mapped onto detector arrays or scanner outputs support high-speed and spectrally resolved FLIM, with pile-up-aware reconvolution fitting permitting rapid sFLIM acquisitions up to 6× faster than classic sFLIM at comparable accuracy (Wahl et al., 2020). More generally, the SPAD-array literature identifies widefield FLIM, multifocal multiphoton FLIM-FRET, SPIM-FCS, time-resolved Raman, and NIROT as core application domains (Bruschini et al., 2019).

Quantum-optical uses rely on event-driven multi-photon capability and spatiotemporal clustering. The intensified Tpx3Cam registers photon pairs from an SPDC source in a dual spectrometer, reaching 0.15 nm wavelength precision and about 3 ns per-pair timing RMS while revealing the expected anti-correlation in signal and idler wavelengths (Nomerotski et al., 2022). The same camera family has been used to count Hong–Ou–Mandel bunched optical photons: clusters from a single spatial mode can be separated spatially and temporally, a clear HOM dip is observed in inter-fiber coincidences, corresponding bunching peaks appear in single-fiber coincidences, and the measured HOM visibility is 42 ± 3% (Nomerotski et al., 2020).

Passive imaging can also be cast in time-resolved single-photon terms. Passive inter-photon imaging measures the time delay between successive detected photons and uses the statistics of these delays as an intensity cue, experimentally demonstrating scenes with a dynamic range of over ten million to one (Ingle et al., 2021). This does not use synchronized illumination, but it remains a time-resolved photon-counting camera modality because it derives radiometric information from timestamp statistics rather than from analog charge accumulation.

6. Performance limits, trade-offs, and development directions

The principal limiting factors recur across architectures: dead time, pile-up, afterpulsing, optical and electrical crosstalk, timing skew, fill factor, and data throughput. The SPAD review reports dead times of 10–100 ns in many CMOS designs, DCR values spanning roughly 0.6 to 100 cps/µm² across processes, afterpulsing from 0.1–10%, and native fill factors from 1–60%, with microlenses and backside illumination used to recover effective sensitivity (Bruschini et al., 2019). Event-driven intensified cameras avoid frame sparsity penalties but inherit intensifier artifacts: the intensified Tpx3Cam reports about 1.4% afterpulsing within 100 ns and about 1.6% within 10 pixels separation, while the broad cluster ToT distribution prevents precise photon counting from amplitude alone (Nomerotski et al., 2022). Timepix-based systems also remain subject to per-pixel dead time, reported as 475 ns + ToT in Timepix3-based instruments (Nomerotski et al., 2022).

Timing electronics impose their own ceilings. The MultiHarp-class architecture reduces electronics dead time to 650 ps per channel, supports more than 80 Mcps sustained aggregate time-tag throughput, and maintains globally consistent timestamps across up to 16 channels, but USB 3.0 bandwidth and host-side overhead still bound sustained rates (Wahl et al., 2020). In compact or portable systems, timing precision can exceed what the rest of the optical geometry can usefully exploit: the 10 ps bin size of the miniaturized NLOS TCSPC module is not the bottleneck when the wall spot size and detector jitter dominate the final resolution (Wu et al., 2024).

Spectral-extension schemes introduce additional system-level trade-offs. The NDTPA EMCCD camera requires a strong, synchronized mid-infrared pump, longer exposures at the single-photon/pixel/pulse regime, and careful spatial and temporal overlap of pump and signal, even though it eliminates phase matching and suppresses pump-only multiphoton background through the long-wavelength pumping condition (Fang et al., 1 Jun 2026). Mid-IR upconversion cameras add conversion-efficiency limits, pump-induced background, and group-velocity-mismatch broadening, while free-space superconducting detectors trade room-temperature simplicity for cryogenic operation and careful optical shielding (Abelson et al., 5 May 2026, Wang et al., 2024).

Current development directions emphasize local processing and scene adaptivity. The reconfigurable SPAD macropixel architecture with LUT-based logic shows that photon counting, direct time-of-flight timestamping, coincidence logic, and programmable weighted sums can be mixed under FPGA-like control directly at pixel or cluster level, with a 17-bit shift register per macropixel and approximately 1.7 ms full-array reconfiguration time for 1024 macropixels at 10 MHz (Milanese et al., 20 Nov 2025). This suggests a move away from fixed-function time-resolved imagers toward cameras that can switch dynamically between D-ToF mode, photon-counting mode, coincidence filtering, and in-sensor event aggregation. A plausible implication is that future time-resolved single-photon counting cameras will be defined less by a single detector technology than by how efficiently they transform photon arrivals into calibrated, application-specific temporal representations at the edge of the sensor.