From measurement point to track
A tracker does not process a perfect flight path. It sees a sequence of uncertain measurements — and must estimate from them how a target is likely to keep moving.
Prior knowledge + measurement = updated knowledge
Bayes does not supply "magical truth". It describes how existing knowledge is changed by new evidence. In tracking, this becomes a repeated cycle across many radar frames.
What do we expect?
The motion so far creates an expectation of where the target is likely to be in the next measurement cycle.
How plausible is the measurement?
The measurement model describes which radar values would be expected for an assumed target state — including measurement uncertainty.
What do we know afterwards?
Prediction and new measurement information are combined. This result becomes the starting point for the next cycle.
Prediction and Measurement Update
In its general form, Bayesian state estimation consists of two steps: first the state is propagated into the future by the motion model, then the distribution is updated with the new measurement.
1. Prediction
xₖ denotes the target state; Z₁:ₖ₋₁ the measurement history so far. The transition model p(xₖ | xₖ₋₁) describes the assumed dynamics.2. Measurement Update
zₖ changes the predicted state distribution via the likelihood p(zₖ | xₖ).The central point: a tracker does not just estimate a single point. It also carries information about how uncertain that estimate is.
More than a position
What the tracker estimates internally and what the sensor directly measures are not necessarily the same quantities.
Cartesian state
A tracker can, for example, estimate position and velocity in Cartesian coordinates. The concrete state definition, however, depends on the method and use case.
Sensor space
Radar can supply range, angle and radial velocity. The mapping between sensor space and state space is therefore often nonlinear.
Where the Kalman filter comes in
The Kalman filter is an efficient recursive state estimator for linear models. It propagates an estimated state and its covariance from measurement cycle to measurement cycle.
Posterior
x̂⁺, P⁺
current state and uncertainty
Prediction
x̂⁻, P⁻
propagated by the motion model
Measurement
z, R
new measurement and assumed measurement uncertainty
Innovation
Deviation between expected and actual measurement
Update
new posterior state and new covariance
In context: For linear Gaussian models, the relevant state distribution can be propagated in closed form via mean and covariance. Extensions such as EKF or UKF exist for nonlinear models. This page treats them only conceptually.
Which measurement belongs to the track?
In clutter, a gate can contain several plausible detections. Then the question "which point is closest?" is often not sufficient.
Why this matters: Clutter, measurement noise, multiple targets and crossing trajectories can make a unique association harder. Gating reduces the candidate space — but does not automatically resolve the association.
And if a frame contains no matching measurement?
A missing detection does not automatically mean the target has disappeared. The tracker can initially continue predicting the state from the motion model alone — but with growing uncertainty.
Short measurement gaps can be bridged. With every unobserved propagation, however, uncertainty should grow; at the same time, a track's credibility can decrease.
Does the track even exist?
State uncertainty and existence uncertainty are two different questions. A tracker can estimate fairly precisely where a track would be — and still be uncertain whether this track corresponds to a real target.
Where is the target likely to be?
Position, velocity and other state quantities are carried as an estimate with uncertainty.
Is this track likely real?
IPDA augments state estimation with a recursively updated existence probability for the track.
What happens when the target suddenly turns?
A single motion model cannot equally well describe straight-flying and strongly maneuvering targets in every situation. IMM handles several motion models in parallel and adjusts their weight based on the measurements.
Constant velocity explains the new direction of motion only insufficiently.
A maneuver model fits the most recent measurements better and therefore receives more weight.
An alternative dynamics model remains as a hypothesis as long as the evidence is not unambiguous.
Important: IMM is not an object-type classifier. It evaluates which motion model currently explains the observed kinematics better.
How the methods fit together
The methods answer different questions within the same tracking chain.
| Method | Core question | Typical role |
|---|---|---|
| Bayes filter | How does a new measurement change our state estimate? | Probabilistic framework for prediction and update |
| Kalman filter | How can state and covariance be estimated efficiently and recursively? | Linear state estimation with explicit uncertainty |
| PDA | Which of several measurements could belong to the track? | Probabilistic data association in clutter |
| IMM | Which motion model explains the current motion? | Weighting several dynamics models in parallel |
| IPDA | How likely is it that the track is real at all? | State estimation plus track existence |
In multi-target scenarios, additional joint association problems arise; methods such as JPDA/JIPDA extend probabilistic association accordingly.
A complete tracking cycle
The real strength does not come from a single algorithm, but from the temporal chain of detection, prediction, association, state update and track management.
Detections in the current frame.
Expected state from the motion model.
Filter out implausible candidates.
Evaluate measurement candidates probabilistically.
Combine prediction and measurement information.
Update dynamics model or model weight.
Propagate the track's credibility.
Provide temporally consistent state information.
Uncertainty is modelled — not eliminated
Bayesian tracking makes uncertainty explicit. How well the estimate performs still depends on models, parametrization, sensor quality and the scene.
Model Error
An unsuitable motion or measurement model can cause systematic deviations.
Incorrectly Assumed Uncertainty
Noise assumptions that are too small can make a tracker overconfident; assumptions that are too large can make the estimate unnecessarily sluggish.
Data Association
Dense clutter and closely spaced targets can make measurement association difficult despite probabilistic methods.
Missed Detections
Tracking can bridge measurement gaps, but cannot compensate for arbitrarily long missing information.
Why this matters for drone detection
Small flying objects are not observed under identical conditions in every frame. Orientation, motion, clutter and viewing geometry can change. Tracking therefore exploits temporal consistency instead of just an isolated measurement point.
Is there a measurement candidate?
Detection looks at the current frame and identifies relevant candidates against background and clutter.
Do measurements belong together over time?
Tracking estimates state, motion and temporal belonging of a moving object.
Which object class is supported?
Classification evaluates features and evidence for an object class. It is not the same as track propagation.
Further reading: Clutter and CFAR · Radar classification with CRNN · Ozense technology · Validation under real site conditions.
Frequently asked questions about Bayesian tracking
Is a Kalman filter a Bayesian filter?
The Kalman filter is a recursive state estimator that fits closely within the Bayesian framework. For linear Gaussian models, the posterior distribution can be propagated in closed form via mean and covariance.
What is a covariance in tracking?
The covariance describes the uncertainty and possible correlation between estimated state quantities. It is not simply "the error", but part of the probabilistic description of the estimate.
What is radar clutter?
Clutter refers to unwanted radar returns not belonging to the target of interest, and the detections arising from them. They can make target detection, and especially measurement association, harder.
What is data association?
Data association describes the decision or probabilistic evaluation of which current measurements could belong to which existing tracks.
What happens when a radar measurement is missing?
The state estimate can initially be propagated with the motion model. Uncertainty typically grows while doing so; depending on track management, the track's credibility or existence probability can also decrease.
What is the difference between tracking and classification?
Tracking estimates state, motion and temporal belonging. Classification evaluates which object class is supported by the observed features.
What does an IMM tracker do?
An Interacting Multiple Model runs several possible motion models in parallel and adjusts their weights based on the measurements. It is not an object-type classification.
Primary sources and further reading
The technical statements are checked against the original publications listed; the sources allow further verification of the methods.
- R. E. Kalman (1960). A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering. DOI:
10.1115/1.3662552 - Y. Bar-Shalom & E. Tse (1975). Tracking in a Cluttered Environment with Probabilistic Data Association. Automatica. DOI:
10.1016/0005-1098(75)90021-7 - H. A. P. Blom & Y. Bar-Shalom (1988). The Interacting Multiple Model Algorithm for Systems with Markovian Switching Coefficients. IEEE Transactions on Automatic Control. DOI:
10.1109/9.1299 - D. Mušicki, R. Evans & S. Stankovic (1994). Integrated Probabilistic Data Association. IEEE Transactions on Automatic Control. DOI:
10.1109/9.293185 - S. J. Julier & J. K. Uhlmann (1997). New Extension of the Kalman Filter to Nonlinear Systems. SPIE Proceedings. DOI:
10.1117/12.280797 - T. Kirubarajan & Y. Bar-Shalom (2004). Probabilistic Data Association Techniques for Target Tracking in Clutter. Proceedings of the IEEE. DOI:
10.1109/JPROC.2003.823149