Most detectors for tiny, dim objects in satellite and infrared video find targets by what changes. They difference consecutive frames, estimate motion, or fit a slowly varying background and flag whatever departs from it. These methods share a known failure. When a vehicle stops at an intersection or an infrared target hovers, the change signal drops to zero, and the background model gradually absorbs the target. The object disappears exactly when it stays put.
Rare-event physics deals with a closely related problem in a different way. Germanium and xenon experiments do not let a candidate event contaminate the baseline it is measured against. They freeze or veto baseline updates during candidate pulses, estimate background from control regions next to the signal region (so-called sidebands), and score candidates with a likelihood ratio built from a known signal shape. This note proposes to bring that discipline to video. Once a target has been seen moving, we stop the background model from learning at its location. We correct that frozen background for slow drift using a ring of surrounding pixels. We then score the stopped target with a matched filter whose template is its own appearance captured while it was moving. Under this design a stationary target is no longer a vanishing signal. It becomes a constant offset whose evidence grows with every frame it stays still.
Core Claim
Motion-based detectors fail on stopped targets because their background estimate is fit to data that contains the target. The fix is not more memory in the detector. It is a change in how the background is estimated. We claim that a "veto-and-sideband" background estimator, paired with a sequential likelihood-ratio test, turns the stationary phase from the hardest regime into one where detection improves with halt duration. The test is a cumulative evidence score that keeps rising while the target stays and falls once it leaves.
Let $B$ be the background at the target's footprint, frozen at the last frame before the target arrived. Let $d_t$ be a drift correction estimated from an annulus of non-vetoed pixels around the footprint. Let $s$ be the target template recorded during the moving phase, and $\sigma^2$ the per-pixel noise variance. The residual is $r_t = I_t - B - d_t$, where $I_t$ is the image patch at frame $t$. The evidence after $T$ stationary frames is
$$\Lambda_T = \sum_{t=1}^{T}\left(\frac{s^\top r_t}{\sigma^2} - \frac{\|s\|^2}{2\sigma^2}\right).$$
$\Lambda_T$ is the log-likelihood ratio of "target present with appearance $s$" against "background only," under Gaussian noise. Its expected value under the target hypothesis grows linearly in $T$, while its noise grows as $\sqrt{T}$. The detection signal-to-noise ratio therefore scales as $\sqrt{T}\,\|s\|/\sigma$. Departure is detected as a change point, using the standard cumulative-sum rule that resets $\Lambda$ when its running slope turns negative.
The prediction is concrete. For targets near or below single-frame visibility, recall during a halt should increase with halt length under our method. Under every motion-based method in the file, it should decrease.
Relation to Prior Work
Several papers motivate the gap by documenting the failure. Motion-inductive Self-supervised Object Discovery in Videos segments objects by clustering optical flow and shows that object boundaries vanish when the flow becomes uniform. Detecting and Tracking Small and Dense Moving Objects in Satellite Videos: A Benchmark builds the VISO benchmark. Its Motion Modeling Baseline combines multi-frame differencing, a low-rank-plus-sparse decomposition, and a five-frame trajectory filter. Vehicles that stop are absorbed into the low-rank background term. Background Subtraction for Online Calibration of Baseline RSS in RF Sensing Networks shows the same absorption in an adaptive radio-signal-strength baseline. Temporal-Emerged Prompting for Segment Anything in Multiframe Infrared Small Target Detection turns motion discrepancies into prompts for a frozen Segment Anything Model (a general-purpose image segmenter). Its authors state that the method goes blind when targets become stationary.
All four treat the background as something learned from the full stream. We instead treat it as a quantity that must be protected from the signal.
Two lines of work come close but address different problems. Implicit Motion Handling for Video Camouflaged Object Detection replaces explicit flow with dense frame-to-frame correlation volumes and long-range transformer attention. Its signal is still frame-to-frame affinity, so it has no positive evidence for a target that looks identical across frames. HTTrack and FATrack keep tracks alive with trajectory priors and search windows centered on predicted positions. Neither measures whether the target is still physically present. They extrapolate or keep a box. That distinction drives the experiments below.
The method itself comes from physics. α-event characterization and rejection in point-contact HPGe detectors separates slow surface-charge release from true bulk events by analyzing pulse tails against a stable baseline. That gives us the principle of decoupling the baseline from candidates. Topological background discrimination in PandaX-III reports results as signal efficiency at a fixed background-rejection level, roughly seven orders of magnitude rejection above 20% efficiency. We adopt that reporting format, because IoU is a poor metric for 4×4-pixel targets.
Enhancing the significance of astrophysical events with multimessenger coincidences raises a subthreshold gravitational-wave trigger from 7.7% to 86% posterior probability when a neutrino coincides with it. It also warns that this boost depends on the assumed source model. We use only a light version of that idea. The observed deceleration of a moving track sets a prior on where a stop occurred. The constraint we take from that paper's warning is that the prior must never be able to create a detection without accumulated pixel evidence.
Research Program
The system has three stages. The first stage is any existing moving-target detector, used unchanged: the Motion Modeling Baseline for satellite video and TEP-SAM for infrared. While a target is moving, its tracked footprint gives the template $s$ and the pre-arrival background patch $B$.
The second stage starts when the track's displacement falls below a threshold. It vetoes background updates inside the footprint plus a guard margin. It then estimates drift $d_t$ from a surrounding annulus using a robust median, which corrects illumination change and sensor gain drift in the same way physics sidebands correct a background level.
The third stage accumulates $\Lambda_T$ and runs the departure change-point test. It reports the stopped target with a confidence derived from $\Lambda_T$, and hands it back to the moving detector when motion resumes.
Two analyses accompany the build. The first separates which part of the gain comes from the veto, which from the sideband correction, and which from the template. The second maps where the method breaks. We expect failure modes when another vehicle stops in the annulus, when lighting changes are spatially nonuniform, and when imperfect frame registration leaves residual motion of the satellite platform. Parallax in the off-nadir view is the hardest case. We will report it as a limitation rather than tune around it.
Proposed Experiments
None of the benchmarks in the file labels halts. We therefore build two evaluation sets.
The first is a real-halt subset of VISO. We will mine the released trajectories for segments with centroid displacement below one pixel over at least $k$ frames, keeping only segments where annotation continues through the stop. Whether VISO annotates vehicles through stops is something we must check first. If it does not, we will hand-label a few hundred halts.
The second is a controlled injection set, following the injection studies used in physics. We take targets cropped from SDM-Car (dim vehicles), NUDT-MIRSDT, and TSIRMT, then reinsert them into real background sequences along recorded trajectories with an inserted stop. The stop's duration, the target's contrast, and the presence of an annulus intruder are varied on a grid. The injection set gives exact ground truth and lets us measure recall as a function of halt duration and signal-to-noise ratio, which no existing benchmark supports.
Matching uses centroid distance (within 2 px for satellite targets) rather than IoU, following the jitter problem noted by Yin et al. The headline metric is halt-phase recall at a fixed false-alarm rate per frame. We also report three secondary measures. The first is departure latency, the number of frames between true departure and the moment the method stops reporting the target. The second is false persistence, the fraction of departed targets still reported $k$ frames later. The third is moving-phase performance, to confirm nothing regresses there: VISO mAP, and F1 on SDM-Car.
Baselines
The naive reference is the moving-target detector alone: the Motion Modeling Baseline on VISO and TEP-SAM on the infrared sets, run on our halt subsets. On full VISO, Yin et al. report the Motion Modeling Baseline at 0.83 mAP, against 0.38 for a Gaussian mixture background model and 0.71 for ClusterNet. TEP-SAM is reported to exceed prior methods by 2.92 IoU on NUDT-MIRSDT and 8.45 IoU on TSIRMT. No reported result was found on Lacuna for any method on halt-only segments, so all halt-phase numbers will be measured by us. We expect these detectors to lose recall within tens of frames. That decay curve is the reference point.
The strongest existing competitor is a trajectory-memory tracker. We use HTTrack, which reports 32.6% success rate and 68.1% precision on VISO, together with a Kalman tracker with track coasting that keeps a box at the last position. These are the systems a practitioner would deploy, and they will score well on halt recall simply by never removing the box. For a fair comparison, every tracker runs on the same registered frames and the same initializing detections, with coasting horizon and thresholds tuned on the same validation split under the same trial budget. Beating them means two things together: matching or exceeding their halt recall, and cutting false persistence after departure by a large margin. On the departure-injection subset, a tracker that never drops a box should score near 100% false persistence. We expect our method to score far lower.
The strongest appearance-only competitor is the CLAHE plus SimAM DSFNet from SDM-Car. It reports an F1 of 82.7 on SDM-Car, compared with 58.5 for plain DSFNet. We will fine-tune it on frames with halts, using the same data.
Ablations isolate each ingredient. The first replaces the local veto with a global slowdown of the background learning rate. The second removes the sideband drift correction. The third replaces the template $s$ with an untemplated energy statistic $\|r_t\|^2$. The fourth drops the deceleration prior. The key ingredient is the local veto. If the global slowdown matches it, the claim fails.
Why not just…?
Why not just keep the track alive with a Kalman filter or HTTrack's trajectory prior? This is the main threat on recall, and we take it seriously. A coasting tracker reports the right location for stops of any length as long as nothing else happens. Its weakness is that it carries no evidence of presence, so it cannot tell a parked vehicle from one that left while undetected. HTTrack's design extrapolates kinematics from five past steps, which predicts drift or continued motion during long stops rather than stillness. The departure-injection experiment settles this. If trackers match our false-persistence rate as well as our recall, the evidence-accumulation idea adds nothing and we will say so.
Why not just detect stopped cars from single frames with a contrast-enhanced detector? For bright vehicles this may suffice, and we expect parity there. For dim targets, the SDM-Car paper shows single-frame static streams confusing vehicles with dark asphalt. The TEP-SAM paper argues that targets at signal-to-noise ratio of 3 or below cannot be separated from noise in a single frame. Our gain should be concentrated in exactly that regime. The contrast-by-duration grid will show where the two curves cross.
Why not just slow or freeze the background update everywhere? A global freeze does prevent absorption. It also fails under illumination change and gain drift, and it leaves the background stale for every other pixel. This is the main ablation. If the gap between a global freeze and a local veto with sideband correction disappears on real VISO halts, the contribution reduces to the template-matched test alone.
Why not just use flow or correlation volumes as in SLT-Net? Frame-to-frame affinity is identical for a stopped target and for the background. Ding et al. and Cheng et al. both show that flow-type cues collapse when texture or motion is absent. We include SLT-Net-style correlation features as an additional baseline on the camouflage-like infrared injections.
Decisive Checks
The idea works if three things hold together. First, halt-phase recall under our method rises or stays flat with halt duration on both the injected and the real VISO halts, while the motion-only baselines decay. Second, at matched recall, false persistence after departure is well below that of the coasting trackers. Third, moving-phase mAP and F1 stay within noise of the unmodified detectors. A separate null check injects no target at all, so that the deceleration prior alone could only produce false alarms. Any detection rate there above the nominal false-alarm rate would mean the prior is inflating confidence, which is the failure Veske et al. warn about.
References
- Motion-inductive Self-supervised Object Discovery in Videos
- Detecting and Tracking Small and Dense Moving Objects in Satellite Videos: A Benchmark
- Background Subtraction for Online Calibration of Baseline RSS in RF Sensing Networks
- Temporal-Emerged Prompting for Segment Anything in Multiframe Infrared Small Target Detection
- Implicit Motion Handling for Video Camouflaged Object Detection
- HTTrack: Learning to Perceive Targets via Historical Trajectories in Satellite Video Tracking
- Foreground-Aware Token Routing Vision Transformer for Real-Time Satellite Video Tracking
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Enhancing the significance of astrophysical events with multimessenger coincidences
- SDM-Car: A Dataset for Small and Dim Moving Vehicles Detection in Satellite Videos