Interactive demonstrations
Explore the handover trigger and the owner–shadow protocol.
The toy example, animated
One target crosses three base stations: it appears in A, brushes through B, and leaves through C — based on the schematic in the paper's appendix. Press play, or drag the timeline; click any numbered event to jump to it.
Table view — the six events
| # | Event | What is sent | Payload |
|---|
With “+ Measurements”, “+ Likelihoods”, or “+ Posteriors”, the co-observation frames between the events additionally carry the per-frame stream counted in the panel above. Payload sizes follow the paper's byte accounting: 196 B per prior, likelihood, or posterior packet, 16 B per measurement; the toy additionally shows illustrative 24 B control messages, which the paper’s reported payload totals exclude.
What to watch for
① Birth. The target first becomes visible to A. A creates the track, assigns local UID 3, and becomes its owner — the label is (A, 3). ② and ③ Prior handover. As the predicted track approaches B's (then C's) region, the handover score \(\Lambda\) crosses the threshold and A sends that neighbor one prior packet. The receiver starts a shadow track — A keeps the original; nothing is relinquished. ④ Shadow pruning. The target leaves B, so B quietly drops its shadow copy and tells the owner. No ownership ceremony is needed — B was only a shadow holder. ⑤ Ownership transfer. When A itself loses sight, ownership moves to the best-scoring shadow holder via a two-message handshake. A retains authority until the acknowledgment arrives. C issues a fresh local UID, and the label changes from (A, 3) to (C, 7). ⑥ Exit. C prunes the track when the target leaves the network.
The four variants share the event protocol. H sends no further evidence. HM sends the selected non-null group to the owner; HL sends a likelihood message; HP sends a full local posterior. HM, HL, and HP return fused posteriors. The toy illustrates traffic and labels, not a numerical implementation of GrBP or its fusion rules.
The deterministic measurement dots illustrate the detection model \(\mu=\rho A_\kappa\delta\), so they thicken near a station and thin at FOV boundaries. The dashed ghost ellipse that appears ahead of the target marks the predicted pose — the density the receiver's shadow prior is initialized from, which is why it reaches the border before the target does.
The trigger: a handover score you can drag
When should the owner speak up? The paper gates every handover on one quantity — the number of detections the receiving station would be expected to make of this track:
The score uses the expected detection count: \(\Lambda\) is the mean count that the receiving station would see, averaged over the predicted density and gated on track confidence. Drag the target below and watch the three factors — the score only clears the threshold when the object is both confident and genuinely entering the receiver's view.