The Gold Standard of Radar | Out-of-Order Observations and the Age of a Track

Measurement time identifies when the recorded physical observation applies. Arrival time identifies when a receiving system gets the corresponding information. Processing time and display time describe later stages. They can be close together, but they are not interchangeable labels.

For a rotating or scanning sensor, different observations may concern different moments within a scan. A frame or file completion time does not automatically identify the acquisition time of every observation it contains.

Now add an estimated state. A tracker may predict a target’s position at a chosen state time using earlier measurements. That state time is neither a claim that a new observation occurred nor necessarily the time of the most recent contributing measurement.

Keep these meanings separate before calculating an age. “Updated at 12:00:03” is incomplete unless the reader knows whether the system received a message, made a measurement or propagated an estimate at that moment.

NIST’s historical Sensors 1588 explanation illustrates generating a timestamp when data is acquired. The page is an archived conceptual resource, not a statement of current radar certification requirements. Its useful principle is that time belongs to the measurement’s meaning.

Contents · Next chapter

For a fictional target, let x = 1,000 metres at time t = 0 seconds. Assume it moves along a defined x-axis at a constant 20 metres per second. Its model position is x(t) = 1,000 + 20t.

At t = 1 second, the position is 1,020 metres. At t = 2 it is 1,040 metres. At t = 3 it is 1,060 metres. These are positions in a supplied reference coordinate, not automatically raw radar ranges.

That distinction matters. A Doppler measurement generally supplies a radial velocity component. It should not be treated as a complete Cartesian velocity vector without the relevant geometry and processing. Our one-axis velocity is a model input selected to isolate the timing question.

Suppose an observation made at t = 2 arrives at t = 2.4. Its 1,040-metre value remains correct for t = 2. At arrival, the constant-velocity model puts the target at 1,048 metres. The eight-metre difference concerns elapsed motion, not necessarily a position-measurement error.

The radar mathematics guide develops coordinate, velocity and uncertainty questions. Here we hold those inputs simple so that the time reference cannot disappear inside a complicated diagram.

Contents · Previous chapter · Next chapter

Consider two messages from that target. Message M1 reports the t = 1 observation at 1,020 metres but arrives at t = 3.0. Message M2 reports the t = 2 observation at 1,040 metres and arrives at t = 2.4.

M2 arrives first even though it describes the later observation. M1 arrives next with an older position. Network, buffering or processing delays can change arrival order without reversing physical motion.

If a display replaces its position with whichever message arrived last, it moves from 1,040 back to 1,020 metres at t = 3. That backward jump is a data-handling effect in this example. It does not establish a target reversal.

Preserve both orderings in the record. Sorting only by arrival time hides the physical sequence; sorting only by measurement time hides how the receiving system actually encountered the information. Each ordering answers a different investigation question.

A useful learning exercise asks for two timelines from the same table. The discrepancy between them is the phenomenon to explain, rather than an inconvenient row to delete.

Contents · Previous chapter · Next chapter

At display time t = 3 seconds, M2’s observation age is 3−2 = one second. M1’s age is 3−1 = two seconds. The message that arrived most recently is the older observation.

If the display simply holds M2’s measured position at 1,040 metres, it lags the model’s current position by twenty metres. If it replaces it with M1’s measured position at 1,020, the difference becomes forty metres.

These differences are not universal radar error values. They come from the stipulated constant velocity and ages. The simple relationship is displacement over the delay = velocity × elapsed time, with direction retained.

The final row is an estimate produced by a motion model. Label it accordingly. A predicted position does not become a newly observed position because it is drawn at the current display time.

The right comparison asks which time the value describes and how it was obtained. An accurate historical position and a useful current prediction can coexist in the same system.

Contents · Previous chapter · Next chapter

Under the exact constant-velocity model, propagate an observation using x(t_ref) = x(t_meas) + v(t_ref−t_meas). The reference time must be named.

For M1 at reference time three, 1,020 + 20(3−1) = 1,060 metres. For M2, 1,040 + 20(3−2) = 1,060 metres. The two estimates agree because the supplied observations and model are consistent.

Now imagine averaging the unaligned values: (1,020 + 1,040)/2 = 1,030 metres. That average mixes positions from different physical times. It is not a justified estimate at t = 3 merely because averaging is a familiar way to combine numbers.

Time alignment is necessary for this comparison, but it is not sufficient for every fusion task. Coordinates, target association, calibration and uncertainty also need to match. Two sources can share a systematic error even when their time labels agree.

The sensor fusion engineering guide develops the wider task of combining observations. The transfer here is specific: decide which physical moment the combined estimate is intended to describe, then account for the age of each input.

Contents · Previous chapter · Next chapter

The previous examples assume the clocks share an accurate time reference. Now give the sensor clock a fixed offset of +0.15 seconds relative to the receiving clock. An observation truly made at t = 1 is stamped 1.15.

When it arrives at true receiving time three, simple subtraction reports an age of 3−1.15 = 1.85 seconds. The actual elapsed time is two seconds. At twenty metres per second, the 0.15-second age error corresponds to three metres of propagation difference in this model.

Delivery latency and clock offset are different. A perfectly synchronised clock does not make a delayed message arrive sooner. A low-latency network does not establish that two timestamps share an accurate reference.

The timestamp’s location in the processing chain matters too. If it is assigned only after a long buffer, it may describe processing rather than acquisition. A beautifully precise timestamp can still label the wrong event.

Record what the time field means, its reference and relevant uncertainty. NIST’s sensor timing material provides background for the acquisition-time distinction; it should not be read as a guarantee about any particular deployed sensor.

For students, repair the mental model before doing more subtraction. Ask which clock produced each number and which event it labels. The difference between two displayed times is meaningful only after those questions are answered.

Contents · Previous chapter · Next chapter

The exact constant-velocity example deliberately has no prediction error. Real estimates have uncertain starting positions, uncertain velocities and possible changes in motion. Extending the prediction farther gives those uncertainties more opportunity to matter.

Use a separate fictional bound exercise. At a correctly aligned measurement time, suppose position error is at most two metres and velocity error at most 0.5 metres per second. Over a two-second horizon, the velocity uncertainty can contribute up to one metre of position error.

Also assume unmodelled acceleration magnitude is bounded by two metres per second squared throughout the interval. Its displacement contribution relative to constant velocity is bounded by 0.5 × 2 × 2² = four metres.

A conservative sum is two + one + four = seven metres under these stated assumptions. This is a worst-case arithmetic bound, not a standard deviation, confidence interval or measured radar accuracy.

If the acceleration bound is unjustified, the seven-metre claim loses its basis. If the clock alignment is also uncertain, that needs an appropriate additional account. Do not quietly add one example’s error number to another example with different conditions.

The learning point is that propagating a state is not merely moving a dot. It carries assumptions about future motion and the reliability of the starting estimate. The model should make those assumptions available for review.

Contents · Previous chapter · Next chapter

An out-of-order observation is not automatically useless. It can help examine a past state, test a proposed association or refine an estimate through an algorithm designed to handle delayed information.

But treating the late value as a new current position is not the same as assimilating it correctly. The tracker needs to preserve the observation’s time and its relationship with earlier processing. Professional tracking methods differ in how they manage this problem.

This article does not prescribe an algorithm for safety-critical systems. The paper example demonstrates the error that must be avoided: arrival at a later moment does not make the physical observation later.

Keep the historical observation and any derived current state distinguishable. A corrected history may be valuable for analysis, while an operational display also needs a clear account of which estimate was actually available at the time.

The air traffic control systems engineering guide connects surveillance timing with a wider service and its assurance. Its professional context adds obligations that a classroom motion exercise cannot establish.

A useful question for any delayed-data system is: what did the message newly tell us about the past, and what does the model now infer about the present? Keeping those sentences separate prevents a great deal of confusion.

Contents · Previous chapter · Next chapter

Begin with a supplied sequence of known measurement times and positions. Process it in time order, then deliver the same observations in a different arrival order. The physical reference trajectory stays fixed.

Inspect the outputs at each receiving time. Does the system distinguish old observations from current estimates? Does it report which time its state describes? Are gaps and predictions visible in the evidence record?

Then introduce separate faults one at a time: a missing message, a duplicated message, a fixed clock offset or an incorrect target identity. Changing all four simultaneously makes it hard to identify which mechanism caused the result.

These are educational test questions. A real validation programme needs task-specific reference truth, operating conditions, uncertainty and acceptance criteria. A smooth path on a screen is not itself evidence that every test has passed.

Preserve the unsuccessful outputs too. Otherwise the demonstration can accidentally become a selection of attractive tracks rather than an account of the system’s behaviour.

Contents · Previous chapter · Next chapter

Sensor A observes the target at t = 1 and reports 1,020 metres. Sensor B observes it at t = 2 and reports 1,040 metres. Their values differ by twenty metres, but both agree exactly with the stipulated motion.

Before declaring a calibration conflict, compare the same physical time. Propagating A’s position to t = 2 with the known twenty-metre-per-second velocity gives 1,040 metres. That resolves the apparent difference within this model.

Now remove the known velocity. The two values alone no longer justify the same exact alignment. We need a motion estimate, additional observations or a suitable uncertainty account. The fact that a correction worked in the first version does not establish that its input is always available.

Change the target to a warming temperature sensor. Here the measured physical quantity itself may respond gradually to the environment. Acquisition time, instrument response lag and network delay become separate questions. Copying a moving-target correction into that setting would ignore the new measurement mechanism.

This is why transfer checks should include a changed condition, rather than only changed names. Ask which part remains common: comparing the same event or state time. Then identify what is new: the physical dynamics, sensor response or information needed to make the comparison.

Contents · Previous chapter · Next chapter

Observation age A position measured at time eight arrives at eleven and is displayed at twelve. Its age at display is four seconds. Delivery delay is three seconds; the extra second concerns later handling or display timing.

Constant-velocity propagation At time eight, x = 400 metres and v = −5 m/s along the stated axis. The predicted position at twelve is 400−5(4) = 380 metres. Keep the velocity’s direction; elapsed time is positive.

Latest arrival or latest observation? One message measured at ten arrives at twelve. Another measured at nine arrives at thirteen. At thirteen, the latter is the newest arrival but the older observation. Store both time fields.

Bound a model change With no initial error, an unmodelled constant acceleration of magnitude one metre per second squared can produce a 4.5-metre displacement difference over three seconds relative to constant velocity: 0.5 × 1 × 3². That is a stipulated model calculation, not a probability statement.

Read the label “Predicted state at twelve using measurements through ten” describes a different evidence state from “measured at twelve”. Both may be useful, provided the distinction remains visible.

Contents · Previous chapter · Next chapter

Does a current timestamp prove current observation? It depends on the timestamp’s definition. It may identify receipt, processing or a predicted state time. Check the acquisition time and contributing data.

Should all late messages be discarded? No universal rule follows from this example. They may support historical analysis or suitable delayed-data processing. Their physical time must remain intact.

Can a higher display refresh rate fix old measurements? Refreshing a screen does not create a new observation. It can display a new prediction, whose quality depends on the model and inputs.

Sources and scope The historical NIST Sensors 1588 page supports the acquisition-timestamp distinction. MIT Lincoln Laboratory’s radar course provides specialist background on tracking and parameter estimation. Every trajectory, message, clock offset and bound here is an original fictional teaching example.

Radar, tracking and combined evidence Return to The Gold Standard of Radar, follow radar mathematics, and continue to sensor fusion or air traffic surveillance systems for their wider applications.

Learning across systems Learning That Transfers into Life explains why distinguishing observation from inference matters. The How X Works hub connects the wider subject library.

When a moving dot seems to disagree with another dot, ask when each position was measured. The answer may reveal a time problem before anyone needs to declare a position problem.

Contents · Previous chapter · Return to Radar