Full sequence re-estimation in each measurement

[adielstatman]

Hi.
I want to demonstrate a multi-target tracking which is not "incremental", i.e. that a new measurement that is input, can affect the former estimations (the whole tracking).
This should be operated since thanks to the new measurement, the former measurements' association probabilities are updated and this should influence, e.g., on the estimation weights.

Since tutorial 8 demonstrates probabilistic data association, what should I update in the code of tutorial 8 in order to get such estimation?
Thanks,
Adiel.

[sdhiscocks]

Smoother may be of interest, but doesn't look at updating association weights.

[orosoman-dstl]

You might want to have a look at multi-frame assignment, implemented in PR #568.

[adielstatman]

Smoother may be of interest, but doesn't look at updating association weights.

Thank you.
When adding in Tutorial 8

from stonesoup.smoother.kalman import ExtendedKalmanSmoother as KS
Smoother=KS(transition_model=transition_model)

and in the end of the tracks loop in the bottom cell

for track in tracks:

...
track=Smoother.smooth(track)

I get:

File "/home/adiel/anaconda3/lib/python3.8/site-packages/stonesoup/smoother/kalman.py", line 78, in _prediction
return state.hypothesis.prediction

AttributeError: 'MultipleHypothesis' object has no attribute 'prediction'

[sdhiscocks]

Looks like regression from previous refactor of smoothers, so they don't function with mixture type tracks. The below should resolve the issue.

diff --git a/stonesoup/smoother/kalman.py b/stonesoup/smoother/kalman.py
index e332ae13..f9ff2221 100644
--- a/stonesoup/smoother/kalman.py
+++ b/stonesoup/smoother/kalman.py
@@ -6,6 +6,7 @@ import numpy as np
 
 from .base import Smoother
 from ..base import Property
+from ..types.multihypothesis import MultipleHypothesis
 from ..types.prediction import Prediction, GaussianStatePrediction
 from ..types.update import Update, GaussianStateUpdate
 from ..models.base import LinearModel
@@ -75,7 +76,17 @@ class KalmanSmoother(Smoother):
         if isinstance(state, GaussianStatePrediction):
             return state
         elif isinstance(state, GaussianStateUpdate):
-            return state.hypothesis.prediction
+            if isinstance(state.hypothesis, MultipleHypothesis):
+                predictions = {hypothesis.prediction for hypothesis in state.hypothesis}
+                if len(predictions) == 1:
+                    # One predictions, this is fine to use.
+                    return predictions.pop()
+                else:
+                    # Multiple predictions, so can't process this.
+                    raise ValueError(
+                        "Track has MultipleHypotheses updates with multiple predictions.")
+            else:
+                return state.hypothesis.prediction
         else:
             raise TypeError("States must be GaussianStatePredictions or GaussianStateUpdates.")

I'll look at opening a PR with this, but you can apply the above for now to test.

[adielstatman]

Great, it works.
Thank you!