Keeping the Faint Ones: Background-Calibrated Coincidence for Detecting Obstacles at 150–250 m

A self-driving truck at highway speed needs to see obstacles 250 m ahead to stop in time. At that distance a car may leave only a handful of laser (LiDAR) returns and a few camera pixels. Today's perception stacks handle this by thresholding each sensor on its own: a camera box scored below 0.3, or a LiDAR cluster with three points, is thrown away as noise before any other sensor or later frame is consulted. Rare-event physics faces a structurally similar problem and solves it differently. Multimessenger astronomers keep "subthreshold" gravitational-wave triggers that are individually unconvincing and promote them when an independent instrument fires at a compatible place and time. They weigh each match against a measured background rate. In Enhancing the significance of astrophysical events with multimessenger coincidences, one trigger's probability of being real rose from 7.7% to 86–97% once a coincident neutrino was folded in.

We propose porting this recipe, rather than just its spirit, to far-field driving perception. Detectors keep every weak camera and LiDAR candidate beyond 100 m. For each range bin and sensor, we measure how often pure background produces a candidate of a given strength. A running log-likelihood ratio then accumulates evidence across sensors and frames until a candidate either crosses a false-alarm budget or decays away. The claim is not that fusion helps at long range, which is known. The claim is narrower: an explicit, range-dependent background model gives the same recall at a guaranteed false-alarm rate as learned fusion. It also transfers to a new LiDAR scan pattern by re-measuring background tables instead of retraining. Across benchmarks, it confirms far objects earlier, measured in metres of warning, than per-sensor thresholding.

Core Claim and Mechanism

Each candidate $c$ carries a posterior log-odds $\ell_t(c)$ of being a real object at frame $t$. Each new observation $x_m$ from sensor $m$ that falls inside the candidate's association region updates it:

$$\ell_t(c) = \ell_{t-1}(c) + \sum_m \log \frac{p_m(x_m \mid \text{object}, r)}{b_m(x_m \mid r)} + \log S_t(c).$$

Here $r$ is range. $p_m$ is the distribution of sensor-$m$ evidence (detector score, point count, return intensity) for true objects at that range. $b_m$ is the same distribution for background, meaning candidates produced where the ground truth says nothing is there. $S_t(c)$ is a spatial overlap term: the integral of the camera's depth distribution along its viewing ray against the LiDAR cluster's position uncertainty, after ego-motion compensation. This plays the role that sky-map overlap plays in the astrophysics paper. The prior $\ell_0$ is the per-range object rate. A candidate is declared once $\ell_t$ exceeds a threshold chosen to meet a target false alarms per kilometre. Because $b_m$ is measured rather than learned end-to-end, the threshold has a direct interpretation, and a lot of calibration signal is available: every unannotated region of every frame supplies background samples. This borrows the physics habit of estimating backgrounds from signal-free sidebands. It also borrows the lesson from the pulse-shape and topology work in α-event characterization and rejection in point-contact HPGe detectors and Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment: suppression factors are only meaningful when quoted with signal efficiency.

The astrophysics paper also carries a constraint. Its posteriors are only as good as the assumed joint-emission model. Our analogue is $p_m(\cdot \mid \text{object}, r)$, which we fit from labelled data per range bin. If that fit is wrong at 200 m, where labels are few, the accumulator will be miscalibrated, so calibration curves are a first-class output.

Relation to Prior Work

Several papers motivate the gap by showing that far-field evidence is too thin for single-shot detection. Point density collapses past 200 m, as documented by Cirrus: A Long-range Bi-pattern LiDAR Dataset and PandaSet. In the latter, PV-RCNN on a 300 m forward LiDAR reaches 49.57 AP for cars at 200 m versus 79.56 within 50 m.

Another group of papers supplies components we reuse but attack the problem inside one network. Far3D turns 2D boxes and depth distributions into 3D queries, but only for boxes that clear a 2D confidence gate. Our camera evidence is exactly its depth distributions, kept below that gate. Super Sparse 3D Object Detection groups sparse points by connected components with cost linear in point count. We use its clusters as LiDAR candidates, but its skeleton sampling is reported to lose small objects, which we measure directly. Self-Supervised Sparse Sensor Fusion for Long Range Perception already attends across warped past frames to 250 m. It is the closest learned alternative, and it offers no explicit false-alarm control.

A final pair constrains our design. An Empirical Analysis of Range for 3D Object Detection shows that far-only training fails, so we keep standard near-and-far detectors and add the accumulator on top rather than training far specialists. It also shows that post-processing is a latency bottleneck, so the accumulator must stay cheap as candidate counts grow. R4D conditions far targets on near reference objects and collapses on empty roads. Our method uses only the target's own multi-sensor, multi-frame evidence, so empty-road scenes are a regime where it should help most.

Research Program

We would build three pieces. First, per-sensor candidate generators with thresholds set very low, drawn from Far3D-style camera outputs and FSD-style LiDAR clusters. Second, background and object evidence tables indexed by range bin (25 m bins from 100 m outward), sensor, and evidence strength, estimated from training splits. Third, the accumulator itself, which associates candidates across frames using ego pose, gates with a radius that grows with range, and applies the update above. The analysis asks three things. How many metres earlier is a far object confirmed at a fixed false-alarm rate? Does the stated false-alarm rate hold on held-out data? And when the LiDAR scan pattern changes, does re-estimating only the background tables recover performance?

Pulse-shape analysis of raw LiDAR waveforms, the most literal physics transfer, is out of scope because none of the listed datasets exposes waveforms. Return intensity is the only "energy proxy" we use.

Baselines

The naive reference point is the unmodified detectors at their standard operating thresholds, evaluated per range bin. This fixes the per-sensor-threshold behaviour we claim to improve on.

The strongest camera-only existing method is Far3D. It reports 24.4% mAP and 18.1% CDS on the Argoverse 2 validation set (Far3D). No range-binned number beyond 150 m was found on Lacuna, so we will re-run it and bin its outputs ourselves. For LiDAR, the Near-Far Ensemble is reported to improve Argoverse 2 CDS by 3.2% over its baseline (An Empirical Analysis of Range for 3D Object Detection). The file gives this only as a relative gain, so we will reproduce the ensemble. For learned temporal fusion, LRS4Fusion reports a 26.6% long-range mAP improvement on its own Long Range Dataset. No Argoverse 2 result for it was found on Lacuna. We will retrain it on Argoverse 2 if code is available, and otherwise use our learned-fusion ablation below as the stand-in.

Two further baselines match our input exactly. The first is lowered thresholds followed by a standard multi-object tracker with a low track-birth threshold, tuned on validation to the same false-alarm budget. The second is a learned fusion classifier: a small gradient-boosted model or MLP trained on exactly the same subthreshold candidate features and history that our accumulator sees. It gets the same tuning budget and training frames as our background tables.

Three ablations isolate the key ingredient. We replace the measured background $b_m$ with raw score products, drop temporal accumulation, and drop the spatial overlap term. To count as beating a baseline, the method must deliver higher recall in the 150–200 m bin at the same false alarms per km, with no increase in near-field false alarms. The difference must survive bootstrap confidence intervals over sequences.

Why not just…?

Why not just lower the thresholds and let a tracker clean up? A tracker with a low birth threshold already accumulates over time, so this is a fair competitor. What it lacks is a range-dependent noise model. A three-point cluster at 60 m is almost surely noise, while at 220 m it may be a truck, and a single birth threshold cannot reflect both. The physics precedent is the MAJORANA and PandaX-III experience: lowering thresholds without a background discriminator floods the analysis. Our false-alarm-matched comparison settles this directly. If the tracker matches us in the 150–200 m bin at equal false alarms per km, the background tables add nothing.

Why not just train a learned fusion model on the same weak candidates? This is the real threat. With enough data, a learned classifier can absorb range-dependent background rates implicitly, and LRS4Fusion shows that learned temporal fusion is strong. We concede that on in-distribution Argoverse 2 data it may tie. The proposal stands or falls on two tests it is less likely to pass. The first is calibration: does the stated false-alarm rate hold on held-out cities? The second is transfer across Cirrus's uniform and Gaussian scan patterns, where Cirrus shows that learned LiDAR models break. There, our method re-measures background tables from a few unlabelled sequences, and the learned model must be fine-tuned. If the learned model wins on both tests, we will report that keeping subthreshold candidates is the whole contribution and the explicit likelihood ratio is not needed.

Why not just add better sensors, such as gated imaging or a foveated long-range LiDAR? Hardware helps. Cross-spectral Gated-RGB Stereo Depth Estimation cuts depth error 39% at 100–220 m with active gated optics. But it needs four synchronised sensors and a pulsed laser, and it does not produce detections. The accumulator is orthogonal to hardware: it applies to whatever candidates a better sensor produces. On Cirrus we test it on top of foveated scanning, where density is already higher, to check that gains persist.

Proposed Experiments and Decisive Checks

The primary benchmark is the Argoverse 2 Sensor Dataset, using ring cameras plus LiDAR evaluated to 200 m. Here we report recall and CDS per range bin at fixed false alarms per km. We also report the safety metric: median first-confirmation distance for objects that eventually enter the 100 m zone, converted to seconds of warning at 75 mph. We check calibration with reliability curves of the accumulated probability per range bin on held-out sequences. Latency must stay within an added 20 ms per frame, comparable to R4D's reported overhead. For generalisation we use Cirrus, training on one scan pattern and testing on the other. There we compare table re-estimation against fine-tuning with the 25% labelled-data regime Cirrus reports as sufficient, and report wDDS alongside the 0.6543→0.6825 range-adaptation reference. As a secondary check, we evaluate PandaSet's forward 300 m LiDAR with its long-focus camera, including scenes with no near-field vehicles, to confirm the empty-road regime where R4D degrades.

The decisive result is a paired comparison against the learned fusion classifier on identical candidates. The idea is validated if our method matches its in-distribution recall and beats it on held-out false-alarm calibration and cross-pattern transfer. If it loses on all three, the background-model hypothesis is wrong.