Counting Trials Before Reacquiring: Look-Elsewhere-Corrected Recovery for UAV Trackers

Drone-mounted trackers lose their target in a predictable way. They search only a small crop around the last known position. When the camera jerks or the target is occluded, the target leaves that crop, and the tracker keeps following whatever patch of background scores highest. The natural fix is to widen the search when confidence drops and to accept a new location only if it looks convincing. That fix contains a statistical trap that particle and cosmic-ray physicists have long accounted for. The highest score in a search region grows with the size of the region even when no target is present. A threshold tuned for a small crop therefore produces more false reacquisitions as the crop grows. Physicists call this the look-elsewhere effect. In the Telescope Array cosmic-ray sky scan, a hotspot of 19 events against 4.49 expected was 5.1σ locally but only 3.4σ once the size of the scanned sky was accounted for (Abbasi et al.).

We propose to make target presence and reacquisition in UAV tracking a calibrated hypothesis test. At every frame, the tracker's own response map (the 2D grid of "how target-like is this location" scores) is compared against a null distribution estimated online from the background part of that same map. The p-value of the best candidate is then corrected for the effective number of independent locations searched, which is set by the window area and the spatial correlation length of the map. The goal is a fixed false-reacquisition rate that holds whatever window size the tracker currently uses. This rate is what makes adaptive search expansion safe. The whole mechanism is a few cheap statistics on a map the tracker already computes, so it can run inside the roughly 27 to 30 frames-per-second budget that DSATrack and MT-Track report on a Jetson AGX Xavier embedded board (DSATrack; MT-Track).

Core Claim

Presence and reacquisition decisions based on a fixed score threshold, or on the classic peak-to-sidelobe ratio, are miscalibrated in a specific way. Their false-alarm rate on target-free frames rises with search-window area. The resulting false reacquisitions, more than missed ones, are what limit adaptive-window recovery. If the threshold is corrected for the number of trials searched, the tracker can grow its search window aggressively after losing the target while holding false reacquisitions at a chosen rate α. This should shorten recovery time on long-term and out-of-view sequences without degrading short-term accuracy.

Concretely, let $r^*$ be the maximum response in a window, let $\hat F_0$ be the upper-tail distribution of background responses fitted from the same frame's map outside a target-sized exclusion zone, and let $N_{\rm eff}$ be the window area divided by the area of one correlation cell of the map. The post-trial p-value is then

$$ p_{\rm post} = 1 - \left(1 - \left[1-\hat F_0(r^*)\right]\right)^{N_{\rm eff}}, $$

and a candidate is accepted when $p_{\rm post} < \alpha$. The form comes from treating the $N_{\rm eff}$ roughly independent cells as separate trials. Background scores are spatially correlated, so we fit a generalized Pareto tail rather than assume Gaussian noise, and we estimate the correlation cell from the map's autocorrelation. Whether this independence approximation is accurate enough is itself one of the things we test.

A second component handles faint targets that never produce a single decisive frame. The source of this component is multimessenger astronomy. There, subthreshold triggers in separate detectors, each unconvincing alone, are combined through a likelihood ratio over spatial overlap and timing. That combination raised one gravitational-wave candidate's probability of astrophysical origin from 7.7% to 86% (Veske et al.). Our analogue treats consecutive frames, and on RGB-thermal data the two cameras, as separate "detectors." Their local p-values are combined only for candidate positions that agree under a loose motion model. The trial count is increased accordingly, to cover the extra freedom of searching over short tracks rather than single points.

Relation to Prior Work

Two lines of UAV tracking work handle spurious peaks deterministically, and they motivate this proposal. ARCF penalizes frame-to-frame changes in the response map with an L2 term (ARCF). DR2Track suppresses the top local maxima as distractors (DR2Track), and its authors report sensitivity to a manually tuned regularization weight. Both change how the filter is learned. Neither asks whether a peak is distinguishable from background at all, and both assume a fixed search region. We leave the host tracker's learning untouched and act only on its decisions.

The modern local trackers are our hosts and supply the motivation directly. DSATrack and MT-Track both name failure to recover once the target exits the crop as their main limitation. MT-Track also gates its memory with a scalar confidence threshold τ. We keep these networks intact and replace τ and the presence decision with the calibrated test.

The Anti-UAV benchmark supplies both the evaluation metric and the hardest regime (Anti-UAV). Its State Accuracy metric rewards correctly reporting absence. It identifies thermal crossover, where the target's temperature matches the background, as the hardest attribute. It also shows the global re-detector SiamRCNN reaching high accuracy at a compute cost unsuited to real time.

From physics we borrow three methods. The trial correction comes from the Telescope Array scan. The continuous-likelihood argument against hard cuts comes from unbinned extended likelihood analyses (J. Lee). The coincidence combination comes from Veske et al.. That last paper also constrains us, because its gains depend heavily on the assumed source model. We therefore keep our motion model deliberately loose and report sensitivity to it.

Research Program

We first run a measurement study to check the premise. For DSATrack, MT-Track, and an ARCF-style correlation filter, we record response maps on target-free crops. These crops come from annotated frames with the target masked out and from Anti-UAV frames labeled absent. We measure how the distribution of the maximum score grows with window area from 1× to 16× the default crop. If the maximum score barely moves with area, the look-elsewhere problem is negligible for learned trackers and the project should stop there. This is a real possibility, since learned heads may produce sparse, peaky maps.

If the premise holds, we build the recovery module. It first runs the calibrated presence test on the normal crop. On failure, it declares the target absent, which matters for State Accuracy, freezes memory updates, and expands the window over successive frames. Where the host tracker allows, it runs the network on tiled crops to cover a larger area. Throughout, it applies the post-trial test and the multi-frame combination to candidate reacquisitions. On RGB-thermal data, the thermal and RGB local p-values are combined only for spatially coincident candidates. The coincidence window is widened to absorb the known registration offset between the unaligned Anti-UAV streams.

We then profile the module's latency on a Jetson AGX Xavier. The statistics themselves are cheap: a tail fit on a subsample of the map plus an autocorrelation. The cost that matters is extra network passes for expanded windows, and this cost is paid only while the target is lost.

Baselines

The naive reference is the unmodified host tracker (DSATrack and MT-Track, using released weights where available) with no absence reporting and no recovery. It anchors how much long-term and out-of-view sequences currently cost.

The strongest direct competitor is the same host with the same window-expansion schedule, deciding presence and reacquisition with a fixed threshold. We run two versions of this threshold. One uses the host's raw peak score, which is what MT-Track's τ does. The other uses the peak-to-sidelobe ratio, the long-standing correlation-filter confidence measure. Its threshold is tuned on a validation split to the same target false-alarm rate our method is set to. The comparison is therefore at matched false-alarm rate in the default-size window, and the question is what happens as the window grows.

A second strong competitor is a learned presence head. This is a small classifier on response-map statistics, trained on Anti-UAV presence labels. It is also given the window size as an input, so that it can in principle learn the trial correction itself.

The accuracy ceiling is a SiamRCNN-style global re-detector. On the Anti-UAV infrared validation set, SiamRCNN is reported at 95.70% precision and 71.52% success (Anti-UAV). We run it on the same sequences to measure the accuracy gap and its compute cost.

For short-term reference points, DSATrack (ViT-B) reports 90.2% precision and 69.4% success on UAV123. MT-Track reports 0.859 precision on DTB70. For UAVTrack112_L, no reported result was found on Lacuna for either host tracker, so we rerun both ourselves.

Three ablations isolate the key ingredients. The first uses local p-values only, with no trial correction, and tests whether the correction matters. The second uses a Gaussian null instead of the fitted tail. The third removes multi-frame and cross-modal combination. All variants share host weights, expansion schedule, validation split, and a tuning budget of one scalar per method.

We count the proposal as beating the fixed-threshold baseline if two things hold. Its empirical false-reacquisition rate must stay within a factor of two of the nominal α across all window sizes, while the fixed threshold's rate grows with area. And it must reach a lower median recovery time and a higher State Accuracy at matched false-alarm rate.

Why not just…?

Why not just use the peak-to-sidelobe ratio with a tuned threshold? This ratio is already a peak-versus-background z-score, so it is our closest relative. Its limitation is that it has no notion of how many places were searched. A threshold that is correct for one window size is wrong for others. The trial-correction ablation and the false-alarm-versus-area curve test exactly this. If the ratio's false-alarm rate turns out flat in window area for these learned maps, our contribution reduces to the multi-frame combination, and we will report that.

Why not just train a presence classifier on labeled absence frames? This is the most serious threat. A classifier given the window size could learn the correction. Our advantages are that we need no absence labels and that we set the error rate explicitly rather than inheriting whatever the training distribution implies. We test transfer directly. The classifier is trained on Anti-UAV and evaluated on UAVTrack112_L and synthetic out-of-view UAV123 sequences, and also at window sizes outside its training range. If it matches us under transfer, the honest conclusion is that the look-elsewhere insight should be used as a training feature rather than as a standalone test.

Why not just run a global re-detector whenever confidence drops? SiamRCNN shows that global search recovers well, but the Anti-UAV authors note its cost prevents real-time execution. Even run only intermittently, it still needs a trigger, and it still searches the largest possible area. That is where uncorrected thresholds are worst. We measure its accuracy and latency on the Jetson and also pair it with our test as a trigger, so the two approaches can be complementary rather than exclusive.

Why not just re-center the crop with a motion predictor? A motion predictor handles smooth motion but not the jitter and occlusion cases where recovery actually fails. We include a constant-velocity re-centering variant of the naive baseline to quantify this.

Proposed Experiments

The decisive calibration check comes first. On target-free crops from UAV123, DTB70, UAV123@10fps, and absent-labeled Anti-UAV frames, we plot the empirical false-reacquisition rate against window area for each decision rule at nominal α of 10⁻², 10⁻³, and 10⁻⁴. We also histogram post-trial p-values, which should be uniform under the null, as in the Monte Carlo calibration of the Telescope Array scan.

For recovery, we evaluate on UAVTrack112_L and on Anti-UAV. We also create synthetic out-of-view events in UAV123 and DTB70 by translating the crop away from the target for known intervals, which gives exact ground-truth absence durations. Metrics are median and 90th-percentile recovery time, false reacquisitions per thousand absent frames, State Accuracy, and standard success AUC and precision. The last two confirm there is no short-term regression against the reported DSATrack and MT-Track numbers.

The thermal crossover subset of Anti-UAV tests whether combining coincident RGB and thermal evidence reacquires targets that neither camera detects alone. We compare against the reported 0.68% mean State Accuracy gain from dual-flow semantic training. We also report sensitivity to the width of the coincidence window. All methods are timed on Jetson AGX Xavier, and a method counts as deployable only if it holds above 25 frames per second on average during loss periods.