Smooth 2D Positions (Kalman Filter)

How it fits

Typical backends

Caveats

• State lifetimeA filter is kept per tracking id inside the running container. An id absent for more than forget_threshold frames is dropped, and reappearing with the same id starts a fresh filter so its position jumps rather than interpolating. While an id is absent but still inside the window the filter coasts on its last velocity (prediction only).
• Parameter interactionSmoothing is set by the ratio of measurement_noise_cov to process_noise_cov: raising the measurement noise (or lowering the process noise) smooths harder but lags real motion, while the opposite passes raw jitter through.
• AccuracyThe emitted direction is the mean smoothed velocity over velocity_history_size frames, so a short window reacts quickly but is noisier than a longer one.
• LimitationEach id is filtered independently with no inter-object motion model, so the smoother cannot reason about interactions or collisions between tracks.
• OrderingThe output preserves the input arrival order of the id-position pairs rather than sorting by id, so consumers should map results back by position.
• State lifetimeAll configuration is captured once at startup; changing any setting requires a redeploy.