Localization Handover Visualizer

Assumptions and limitations

• Reduced to 1D along-track motion at 1 m/s through a door at 40 m so the mechanism is visible. Real transitions have 3D pose, orientation, map-frame geometry and multiple estimators; the structure here is the point, the numbers are not predictions.
• Global pose g = odometry l + offset o. Odometry (LiDAR-inertial) is continuous and drifts as a random walk. GNSS and prior-map localisation each measure the offset. The offset filter predicts with the drift variance and updates with each source.
• GNSS noise and degradation rise from 12 m before the door to 8 m after it; multipath bias rises with them. “Honesty” sets how much of the degradation the receiver reports: at 0 it keeps claiming good accuracy while biased, which is the dangerous case. The indoor map becomes usable 18 m before the door and fully confident 8 m before it.
• The map is registered with a constant error relative to the GNSS frame. That error is a main source of pose jumps: whichever source dominates indoors defines the frame the pose ends up in.
• S1 uses a GNSS-only offset filter, then a map-only filter, switched when the indoor map is confident or GNSS reports σ above 2 m. S2 averages those two filters by their variances without aligning the frames. S3 is one filter fed by both sources; options add 3σ innovation gating, estimation of the map-to-GNSS alignment while both are healthy (reported σ below 1.5 m), and a slew-rate limit on how fast the published pose may absorb a correction.
• The alignment estimate is only as good as the GNSS during the overlap; with a biased GNSS in that window it learns the wrong offset. Gating and a GNSS quality check are what protect it.
• No loop closure, no map update, no failure of the indoor localiser. A real design also needs recovery when the indoor localiser fails; see Module 12.