Anchor-Level Blockage States for UWB Tracking Under Obstruction

Ultra-wideband (UWB) radios locate a tag by timing nanosecond-scale pulses between it and fixed anchors. In an empty room they are accurate to about 10 cm. When a wall, a metal cabinet, or a passing person blocks the direct path, the pulse arrives late, and the tracker's position jumps by decimeters to meters. Current systems handle this in one of two ways. Some classify each measurement as clear or blocked on its own and throw out the blocked ones. Others feed every measurement into a Kalman filter that assumes Gaussian noise. We propose a third approach. Blockage is a property of an anchor at a point in time, not of an individual measurement. Many measurements are functions of the same few anchors, so the right output to predict is a small set of per-anchor blockage labels, inferred jointly across anchors and across time. The tracker then uses those inferred states to weight or correct the ranges instead of discarding them.

The observation that makes this more than smoothing is specific to time-difference-of-arrival (TDOA) UWB, where each measurement is the range difference between a pair of anchors. If anchor $i$ is blocked, every pair that contains $i$ gets a biased value with a predictable sign. The label on a measurement therefore breaks down into the labels of its two anchors. We make this breakdown explicit as a hierarchical label decomposition, and we decode it with a conditional random field (CRF). A CRF is a model that scores whole configurations of labels jointly rather than one label at a time. Here it ties anchors that share measurements and ties each anchor to its own state at the previous step.

Core Claim

We claim three things. First, predicting factorized anchor-level blockage states improves tracking accuracy in cluttered UWB environments beyond per-measurement classification and beyond per-measurement temporal smoothing, when the output is a factorization into visibility, obstruction cause, and static-versus-dynamic, and the states are jointly decoded by a CRF with cross-anchor and temporal potentials. Second, the gain comes mostly from the cross-anchor coupling, which exploits the fact that pair measurements share anchors. Third, the factorization lets different datasets supervise different parts of the label. Some recordings label obstruction material but give no trajectory. Others give trajectories whose blockage labels can only be derived from residuals against ground truth.

Where This Sits in Prior Work

Four lines of work shape the design, and none of them occupies it.

Hierarchical label decomposition in indoor positioning so far means spatial hierarchy over Wi-Fi fingerprints. Kim et al. replace 905 flat location classes on UJIIndoorLoc with parallel building, floor, and location heads (118 outputs), reaching a 9.29 m mean error. Cheng et al. tune the same design with particle swarm optimization and report 9.42 m. These papers motivate factorizing the output. They are also a warning about what not to factorize. Building, floor, and zone are the wrong target for a centimeter-level technology, because UWB geometry already resolves them, and illegal floor-and-building combinations can be removed by a lookup mask. We keep their idea, that a factorized output shares statistical strength across combinations, and apply it to the variable UWB actually gets wrong: the propagation state of each anchor.

The UWB literature supplies the physics and the failure modes. Gezici frames UWB positioning as two steps: estimate ranging parameters, then solve for position. Blockage by an obstacle, called non-line-of-sight (NLOS) propagation, corrupts the first step with positive bias. Maranò et al. classify each link with a support vector machine on waveform features (energy, rise time, kurtosis) and reach 91% accuracy on MIT measurements. They also show that discarding NLOS links causes outages once fewer than three anchors remain, and that soft ranking of anchors avoids this. Their classifier looks at one snapshot and one link at a time. We keep their soft-use principle, and replace the per-link, per-snapshot decision with a joint posterior over all anchors. UTIL supplies the benchmark and the motivating negative result. An error-state Kalman filter reaches about 10 cm in obstacle-free constellations, but "deteriorated significantly" in the cluttered constellation #4, where static obstacles cause consistent biases and dynamic ones cause less predictable errors. Locate3D (Fast Localization and Tracking in City-Scale UWB Networks) shows that adding geometric constraints helps. It does so through continuous optimization, not discrete state inference, and we do not compete with it on network scale.

CRF work supplies the method. The linear-chain CRF with neural emissions, used for example by Jagannatha to rule out illegal tag sequences, gives us temporal potentials. The factorial CRF in ImageSpirit couples separate label layers (objects and attributes) with mean-field inference at interactive rates. That is the closest structural precedent for coupling our factor layers, though it works on static image pixels. Spot is the only CRF-based indoor localizer in the file. It uses device-free Wi-Fi with a spatial CRF plus a hidden Markov model and reports under 1.1 m median error. Its CRF sits over entity positions, not over the reliability of the sensors. We move the random field from "where is the person" to "which anchors can be trusted right now."

Research Program

At each timestep $t$, each anchor $i$ gets a latent label $z_{i,t} = (v, c, m)$. Here $v$ is the visibility (clear or blocked), $c$ is the obstruction cause (none, or one of a small set of material classes), and $m$ says whether the blockage is static or moving. The label on the TDOA measurement between anchors $i$ and $j$ is not predicted directly. It is determined by $(z_{i,t}, z_{j,t})$, and so is its expected bias: approximately the bias of $i$ minus the bias of $j$. This deterministic link from anchor labels to measurement labels is the hierarchical decomposition.

The CRF energy has three parts. Unary potentials come from a small neural encoder over per-measurement observables: the signal-to-noise ratio and power-difference values UTIL records, plus the innovation of each measurement against the filter's prediction. Pairwise temporal potentials connect $z_{i,t-1}$ to $z_{i,t}$ and are learned separately for static and moving blockage, since the two persist differently. Measurement-consistency potentials tie each pair's observed innovation to the bias its two anchor labels imply. With eight anchors, the joint label space per timestep has a few thousand configurations. A flat classifier over it would be starved of data, which is why we factorize.

Inference runs in two modes. The first is online forward mean-field, in the spirit of ImageSpirit, for real-time use. The second is fixed-lag smoothing for offline analysis. The anchor marginals then enter the error-state Kalman filter as a Gaussian-mixture likelihood per measurement, with learned material-specific bias means. This means nothing is ever discarded: an anchor with a 0.9 blockage probability is down-weighted and bias-corrected, not removed.

One circularity needs care. Innovations depend on the filter state, which depends on the labels. We handle this by alternating one filter prediction, one CRF update, and one filter correction per step. We will report whether this loop ever locks into a wrong fixed point.

Training uses partial supervision. The UTIL identification recordings give material and visibility labels for known links but no trajectories, so they supervise $c$ and $v$. The UTIL flight recordings, combined with Vicon ground truth and surveyed anchor positions, give true TDOA values, so residuals yield $v$ and approximate bias magnitudes. Segments with dynamic obstacles give $m$. The factorized loss masks whichever factors a given sample does not label. A known limitation is that UTIL's hardware (DWM1000) exposes only summary signal statistics, not the full waveforms Maranò et al. exploited. The waveform-feature advantage reported on the MIT data may therefore not transfer, and we test this rather than assume it.

Baselines

The naive reference point is the error-state Kalman filter shipped with UTIL, run on all measurements with its default Gaussian noise model. It is the method UTIL reports at roughly 10 cm in obstacle-free constellations. On constellation #4, UTIL states only that performance "deteriorated significantly" and gives no figure in the file. We will therefore re-run it ourselves and report its error on the same held-out flights.

The strongest existing method for this question is per-measurement NLOS classification with soft use, following Maranò et al. We train a classifier on the same observables and the same labeled data as our model. At the reported level, it reached 91% LOS/NLOS accuracy against 84% for a parametric baseline on MIT data. Its scores then set per-measurement noise inflation in the same filter, and we also run the top-ranked-anchor selection rule. This is the right competitor because it already avoids the outage failure and differs from our model only in lacking joint structure. We run it in two variants: soft-ranked, and hard-discarding, the latter to reproduce the outage result on UTIL.

A standard robust-filter baseline is the same filter with chi-square innovation gating and a Huber-weighted update, which needs no learning at all.

The ablations that isolate the key ingredient are a per-measurement hidden Markov model without anchor sharing (temporal only), our CRF without temporal potentials (anchor sharing only), and a flat classifier over joint anchor configurations in place of the factorized heads. Every learned model gets the same encoder size, the same training flights, and the same hyperparameter search budget, measured in trials on a validation flight.

We will say the proposal beats the strongest baseline if two conditions hold. Position RMSE against Vicon must be lower across held-out constellation-#4 flights, by a margin larger than the flight-to-flight spread under paired bootstrap resampling. And the improvement must be concentrated in multi-anchor blockage segments, where the structural argument predicts it.

Why not just…?

Why not just put a CRF over building, floor, and zone heads on UWB, mirroring the Wi-Fi work? Because in UWB the hierarchy is not where the error lives. Kim et al. needed factorized heads because Wi-Fi fingerprints are ambiguous at meter scale. UWB geometry resolves floors directly, and any illegal combination can be removed by a lookup mask with no learning. A CRF there would add cost and solve a problem nobody has.

Why not just smooth each measurement over time with an independent hidden Markov model? This is the most serious threat. Static obstructions persist, and a per-pair temporal model may capture most of the benefit. Our argument is that TDOA pairs share anchors, so one blocked anchor corrupts several measurements at once, and evidence from all of them should be pooled. The temporal-only ablation settles the question directly. If it matches the full model, the cross-anchor coupling is not worth its cost, and we will report that result.

Why not just gate outliers in the Kalman filter? Gating rejects large innovations, which reintroduces the discard problem Maranò et al. documented: when several anchors are blocked, too few survive. Gating also cannot catch moderate, consistent static biases that look like plausible measurements. The robust-filter baseline measures how much of the gap this closes without learning.

Why not just regress each measurement's bias with a neural network and subtract it? Direct bias regression is a real alternative, but it has no notion of persistence or of shared cause across pairs. It also cannot use material labels from recordings that lack trajectories. We include it as a baseline with the same encoder, and compare errors separately on dynamic-obstacle segments, where UTIL reports errors are least predictable and a regressor would be expected to fail.

Proposed Experiments

The main experiment is leave-one-flight-out evaluation on UTIL constellation #4. We report position RMSE and 95th-percentile error against Vicon, outage rate (the fraction of timesteps with fewer than the minimum number of usable anchors), and per-timestep compute cost. Results are broken down by segment type: clear, single-anchor blockage, multi-anchor blockage, and dynamic blockage.

A second experiment measures anchor-state classification accuracy on the UTIL identification recordings. For material-cause accuracy we train with and without the flight data, to test whether the partial supervision enabled by factorization actually transfers between the two kinds of data. A third experiment tests the no-harm condition: running the full model on obstacle-free constellations #1 to #3 must not degrade the roughly 10 cm baseline.

The decisive check is the multi-anchor blockage slice. If the full model beats the per-measurement hidden Markov model there, but not on single-anchor segments, then the anchor-level decomposition is doing the work we claim. If the two tie everywhere, the central idea is wrong, and the contribution reduces to temporal smoothing of the soft-ranking approach.