ToF LiDAR signal processing: pulse timing, peak detection, and range aliasing

Pulsed time-of-flight LiDAR measures range by emitting discrete laser pulses and precisely timing the round-trip delay of the return signal — making accurate peak detection the single most critical signal processing challenge in any ToF architecture. The hardware required to meet that challenge is substantial: a time-digital converter (TDC), a multi-pixel photon counter (MPPC), and an analog-to-digital converter (ADC) operating at 250–500 MHz, all coordinated to extract the amplitude and precise timestamp of each reflected pulse.

Baidu USA LLC has filed extensively across this peak detection problem. Their primary approach pairs a TDC with an MPPC: the TDC’s configurable trigger threshold indicates when sufficient MPPC pixels have fired, triggering the peak detection module to sample and store photon counts. Peak amplitude is derived from the highest sample count; peak timing is determined as the midpoint of the TDC trigger window. A complementary Baidu approach employs spline interpolation to reconstruct the continuous analog waveform from sparse digital samples at 250–500 MHz, enabling sub-sample peak localisation — with multiple laser emitters time-multiplexed onto a single ADC channel.

A third Baidu hardware variant pairs a single-photon avalanche diode (SPAD) detector with an avalanche photodiode (APD) to simultaneously capture intensity and ToF data from the same return pulse, broadening dynamic range and enabling amplitude-calibrated point clouds. Samsung Electronics advances the pipeline further through cross-correlation signal correction: their method identifies the main receiving signal (maximum value at a given time point) and uses it to correct sub-receiving signals corresponding to side-lobe artifacts or multi-target returns. The correction factor is defined as the ratio of side-lobe intensity to main-lobe intensity of the transmitted pulse — enabling multi-echo ToF estimation critical for vegetated or semi-transparent obstacle detection.

LG Innotek has pursued two further approaches to ToF amplitude estimation. A differential comparator design feeds the raw analog LiDAR output and a time-delayed version into a comparator: the rising edge marks range-onset and the pulse width encodes reflected signal amplitude, enabling simultaneous range and intensity from a single comparator circuit. Separately, LG Innotek’s compressed sensing approach maps sparse ToF values to frequency domain sinusoids using numerically controlled oscillators, reconstructing the signal from far fewer samples than a direct histogram would require — dramatically reducing memory and processing overhead for long-range LiDAR.

Range aliasing is the most consequential ToF-specific artifact. Waymo LLC addresses it through two complementary strategies: extended detection periods that identify returns falling outside the standard window, and time-varying pulse dither — deliberate timing jitter in the emission sequence — so that the same object yields return-time signatures consistent with only one of the hypothesised ranges across multiple pulses. According to WIPO patent databases, Waymo has pursued this protection across EP and IL jurisdictions in addition to the US, reflecting the global importance of long-range ToF reliability for autonomous vehicle deployment.

FMCW LiDAR: coherent beat-frequency analysis and instantaneous velocity extraction

FMCW LiDAR transmits a frequency-modulated continuous wave rather than discrete pulses, and detects range and velocity from the beat frequency generated by mixing the return signal with a local oscillator copy of the transmitted signal — extracting both measurements simultaneously from a single sensing event. Because the Doppler shift appears directly in the beat frequency, radial velocity is an algebraic output of the coherent mixing process itself, not a derived quantity requiring temporal differencing.

“FMCW LiDAR resolves the instantaneous velocity of an object during a single sensing period without reference to future sensing periods — enabling perception architectures that ToF systems cannot replicate without multi-frame accumulation.”

Aurora Innovation’s patent portfolio documents this capability in detail. Their phase-coherent LiDAR component generates data that simultaneously indicates each point’s range and velocity, enabling the instantaneous velocity of an object to be resolved during a single sensing period without reference to future sensing periods. The perception subsystem then processes instantaneous range and velocity jointly for each point in the point cloud, passing both object pose and velocity to the planning subsystem in a single frame — architecturally distinct from ToF, where velocity must be inferred by differencing range measurements across multiple scan frames.

This single-frame velocity capability has direct operational applications. Aurora Operations applies FMCW LiDAR returns from an attached trailer to derive the trailer’s yaw rate per sensing cycle, enabling real-time autonomous control adaptation — a use case that would require multi-frame accumulation under a ToF architecture.

FMCW systems face their own unique signal processing challenges, however. Because coherent detection is used, partial occlusion of the transmit/receive optical window introduces a distinctive beat-frequency artifact that differs from normal background and range-related beat tones. Aeva Incorporated has patented systems specifically for detecting these window obstruction artifacts in FMCW LiDAR, analysing their operational impact on detection performance, and implementing mitigation strategies. This is fundamentally different from ToF window occlusion, which manifests primarily as amplitude attenuation or point-cloud density reduction — making the detection and mitigation algorithms for each modality incompatible. Standards bodies such as IEEE have begun addressing coherent LiDAR interference standards, and automotive safety frameworks from ISO are increasingly relevant to both modalities’ fault detection requirements.

Velocity estimation, calibration pipelines, and sensor fusion

Velocity estimation: inferred vs. native

ToF-based velocity estimation requires multi-pulse or multi-frame processing because the modality provides no direct Doppler measurement. PlusAI describes a method where two successive ToF pulses are emitted with a precisely configured inter-pulse threshold interval; by comparing the ToF of the first and second return signals from the same object in motion, the system derives an object velocity estimate. This two-pulse method introduces latency and is susceptible to frame-to-frame data association errors in crowded scenes. Volvo’s deep-learning approach uses convolutional neural networks to estimate ego-velocity from sequential ToF frames by estimating rotation and translation between frames, combining that with per-range-band relative velocity estimation to derive target absolute velocity.

FMCW, by contrast, delivers instantaneous per-point Doppler velocity natively, eliminating this multi-frame dependency. This difference is not merely a latency concern — it reduces the architectural complexity of the entire downstream perception stack, as velocity becomes a first-class point-cloud attribute rather than a derived quantity computed after object detection and tracking.

Calibration pipelines

ToF LiDAR produces geometrically dense point clouds that require careful coordinate-system calibration to maintain accuracy over time and distance. Baidu USA LLC has developed two parallel calibration strategies: cross-validation-based calibration, which iteratively adjusts coordinate converter parameters to minimise obstacle location inconsistency across scan frames; and static reflection map calibration, which compares a dynamic LiDAR reflection map against a pre-collected reference and optimises the converter to maximise overlap. Aurora Operations addresses longitudinal bias in ToF scan alignment by classifying each LiDAR point as “explained” or “unexplained” relative to a surfel map, then iteratively shifting the scan to minimise unexplained points. Notably, this Aurora patent (2026) demonstrates that even FMCW-equipped AV operators must solve analogous geometric alignment problems — calibration complexity is architecture-independent, but solution strategies differ.

Sensor fusion and dynamic thresholding

ToF point clouds are routinely fused with camera data in AV perception pipelines. Baidu USA LLC’s flexible synchronisation patent describes a hardware-level control architecture that dynamically adjusts LiDAR and camera control signals based on detected features — such as missing data or synchronisation drift — ensuring the two data streams remain temporally aligned for fusion. LG Innotek’s dynamic detection threshold patent applies specifically to pulsed ToF systems: a comparator receives the analog output from a light detector and generates a digital signal based on a tunable threshold; the controller iterates through threshold adjustments and aggregates the resulting digital outputs to improve detection robustness in varying ambient conditions.

FMCW receivers, operating on continuous-wave coherent signals, apply noise mitigation differently — through coherent integration gain and adaptive windowing of the beat spectrum. The output point-cloud type changes the downstream fusion architecture: a purely geometric ToF point cloud fuses differently with camera data than an FMCW point cloud that carries per-point velocity. Research published by Nature on photonic integrated circuits for LiDAR underscores the growing interest in coherent receiver miniaturisation that will shape next-generation FMCW fusion architectures.

Head-to-head: where the two architectures diverge for AV perception

The key architectural divergence between ToF and FMCW is in the velocity processing pipeline. ToF systems expend significant signal processing resources on accurately timestamping return pulses and then inferring kinematic state from temporal differences. FMCW systems derive velocity as an algebraic output of the coherent mixing process itself, enabling Aurora Innovation’s perception layer to skip multi-frame tracking for velocity and directly feed instantaneous velocity into the planning subsystem.

Innoviz Technologies’ adaptive noise mitigation and dynamic light flux patents — which adjust pixel-by-pixel sensor sensitivity based on scene content — are principally designed for pulsed-ToF flash and scanning systems, illustrating that field-of-view management trade-offs are handled at the hardware level in ToF architectures. Both modalities confront these trade-offs, but the solution strategies are modality-specific.

Patent landscape: who owns what in LiDAR signal processing

The approximately 60-patent dataset spans active, pending, and inactive filings across US, CN, JP, KR, EP, WO, DE, and IL jurisdictions, with dominant assignees splitting clearly along modality lines. Baidu USA LLC is the most prolific assignee by volume, with the dominant focus on ToF LiDAR signal processing infrastructure: peak detection (TDC/MPPC and spline-based), SPAD/APD hybrid receivers, and multiple calibration paradigms. Their CN-filed counterparts reflect systematic global prosecution of the same families, indicating a strategy of protecting the entire ToF signal chain for autonomous driving vehicle applications.

Waymo LLC concentrates on range aliasing resilience in ToF systems, with multiple patent families across EP and IL jurisdictions covering both extended-period detection windows and multiple-hypothesis dither-based pulse disambiguation. Aurora Innovation and Aurora Operations represent the primary FMCW/phase-coherent LiDAR patent presence, spanning object classification, yaw parameter estimation, and bias correction. Aeva Incorporated holds the FMCW-specific window occlusion detection patents, positioning them as the primary IP holder for FMCW fault detection in vehicular environments. LG Innotek is diversifying across differential comparator amplitude estimation, compressed sensing for sparse ToF histograms, and dynamic detection threshold control — reflecting a component-supplier strategy to serve multiple AV OEMs.

According to EPO patent trend data, automotive sensing and perception technologies have seen sustained growth in filings over the past five years, consistent with the multi-jurisdictional prosecution strategies visible in this dataset. The PatSnap IP Intelligence platform provides deeper cross-jurisdictional family analysis for teams tracking this space.