Find your next learning step
A reference window may contain ordinary clutter, strong interfering targets and a clutter edge at the same time. Trimmed-mean CFAR makes the censoring choice visible instead of letting every cell vote equally.
Open the chapter contents · Radar root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The teaching chapters remain open; these grouped routes form the index.
Route 1 · See the estimator
Route 2 · Work the window
Route 3 · Stress the scene
Route 4 · Check limits and evidence
Route 5 · Transfer and continue
CHAPTER 1 OF 12 · SEE THE ESTIMATOR
1. A reference window should not let every cell vote equally
Cell-averaging constant false alarm rate detection estimates local clutter power by averaging reference cells around a cell under test. It is efficient when the reference window is homogeneous. It can fail badly when a few reference cells contain interfering targets or when the window crosses a clutter edge.
Trimmed-mean CFAR begins by sorting the reference powers. It discards a chosen number of the smallest and largest samples, then averages what remains. The upper trim is often used to reduce target contamination; the lower trim can change behaviour at clutter transitions. The method sits between cell averaging and a single order statistic.
This article extends the Radar root, CFAR reference-window contamination and order-statistic CFAR. The new design variable is not only rank. It is how much of the sorted window remains in the average.
CHAPTER 2 OF 12 · SEE THE ESTIMATOR
2. Trimming depth defines the noise estimator
Let N reference powers be sorted as Z_(1) no greater than … no greater than Z_(N). Choose L low samples and H high samples to discard. The retained count is M = N – L – H, and a simple trimmed estimator is the average of Z_(L+1) through Z_(N-H).
Three familiar detectors appear as limiting cases. With L = H = 0, the estimator is the cell average. If only one ordered value is retained, it behaves like an order-statistic estimator. Between those cases, several ordered cells contribute, so one unusual retained sample has less leverage than it would in a one-cell statistic.
The threshold multiplier must be calibrated for N, L, H, the assumed clutter law and any correlation or integration. Reusing the multiplier from cell averaging is a false-alarm error, even if the numerical trimmed mean looks plausible.
CHAPTER 3 OF 12 · WORK THE WINDOW
3. Worked example: sixteen sorted reference cells
Consider this original reference-window ledger in normalized power units. Two large values represent strong interference; smaller values could be ordinary fluctuation or the low side of a clutter edge.
| Ordered position | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Power | 0.2 | 0.3 | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | 1.0 | 1.1 | 1.2 | 1.3 | 1.4 | 1.5 | 1.7 | 8.5 | 12.0 |
The ordinary mean is 35.7/16 = 2.231. Removing only the two largest cells gives 15.2/14 = 1.086. Removing the two smallest and four largest leaves the ten values 0.5 through 1.4, whose mean is 9.5/10 = 0.950.
The estimator has moved because the design says which evidence is allowed to define clutter. That is useful only when the trimming rule matches the scene and the threshold multiplier is recalibrated.
CHAPTER 4 OF 12 · WORK THE WINDOW
4. Censoring high cells protects against a finite number of interferers
If at most H reference cells are contaminated by strong targets and those cells occupy the largest ranks, an upper trim of H can remove their direct influence. If H+1 strong cells enter the window, at least one remains. This is a finite robustness budget, not general immunity.
| Low trim L | High trim H | Retained cells M | Estimate in worked window | Main risk |
|---|---|---|---|---|
| 0 | 0 | 16 | 2.231 | Strong masking from two interferers |
| 0 | 2 | 14 | 1.086 | Low-edge samples still pull downward |
| 2 | 4 | 10 | 0.950 | More censoring, less effective sample support |
| 4 | 4 | 8 | 1.050 | Genuine distribution tails may be removed |
Interference spillover matters because target energy may occupy several adjacent range–Doppler cells. One physical target can consume more than one unit of the trimming budget. Count contaminated cells, not only distinct objects.
CHAPTER 5 OF 12 · WORK THE WINDOW
5. Calibration must follow the trimmed statistic
For exponentially distributed independent clutter, the distribution of the trimmed sum can be derived from exponential order-statistic spacings. For correlated, integrated or non-exponential clutter, numerical calibration or a validated model may be more practical. In every case, calibration must target the chosen false-alarm probability with the exact window geometry and trimming rule.
Ozgunes, Gandhi and Kassam introduced a variably trimmed mean detector whose retained set changes by a data-dependent rule. Their work shows that trimming can improve behaviour at clutter transitions and in multiple-target scenes, but the adaptive rule becomes part of the detector and must be included in calibration.
A Monte Carlo estimate of a very small false-alarm probability needs enough independent trials and a confidence interval. Reporting zero false alarms from too few trials is not evidence that the rate is zero.
CHAPTER 6 OF 12 · STRESS THE SCENE
6. Changed context: the clutter edge is genuine, not contamination
Suppose the leading half of the window samples low clutter and the lagging half samples high clutter. Sorting erases which side each value came from. Upper trimming may discard the high-clutter half as if it were interference and set a threshold that is too low for a cell under test on the high side.
That is why spatial order and rank order answer different questions. A greatest-of or smallest-of split-window rule, a switching detector or explicit edge classification may be more appropriate when side information matters. Trimmed mean is not automatically an edge detector.
The transfer check is to retain the original cell positions alongside the sorted ledger. If the cells selected for trimming cluster on one physical side, the estimator may be diagnosing a transition rather than isolated targets.
CHAPTER 7 OF 12 · STRESS THE SCENE
7. Changed context: interference is moderate and spread out
Strong isolated interferers usually rise to the top ranks. Moderate interference spread across many cells may not. It can raise much of the retained set and inflate the threshold even after the planned upper trim. Conversely, aggressive trimming can lower the effective sample size and increase threshold variability.
Test spillover patterns, not only single-cell spikes: one broad target, two separated targets, a Doppler skirt, a range sidelobe pattern and a clutter edge. Use the same guard-cell geometry that the deployed processor uses.
This changed context connects to compound-Gaussian clutter and CFAR false-alarm inflation. Heavy-tailed texture can create high ranks without a discrete interfering target, so a fixed trim has a different meaning.
CHAPTER 8 OF 12 · STRESS THE SCENE
8. Correlation reduces the independent evidence left after trimming
Reference cells are often correlated because of filtering, oversampling, coherent processing or spatial structure. Ten retained cells do not necessarily provide ten independent observations. Correlation can alter both the mean-estimator distribution and the false-alarm multiplier.
Measure or model the correlation after the same processing chain used in deployment. If a calibration assumes independent exponential samples but the live window is correlated, “constant” false alarm becomes only a label.
The same caution applies to noncoherent pulse integration. The shape of the clutter distribution changes with the effective pulse count, and correlation can make that count smaller than the nominal number of pulses. Continue with noncoherent integration and correlation loss.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and misconceptions
Trimming is not free robustness. It removes data, changes the estimator distribution and can increase detection loss in homogeneous clutter. A rule tuned to remove four high cells assumes those ranks are expendable; if they are genuine clutter evidence, the estimate may be biased downward.
The trim count is not the number of physical targets. One target may contaminate several cells, and several weak targets may remain below the cut. Sorting also discards position unless a separate record is retained.
Finally, the estimator and multiplier are one design. Changing L or H without recalibrating the multiplier does not preserve the false-alarm rate. Detection probability should be compared at matched false-alarm probability, not at matched multiplier.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Evidence and verification route
The foundational reference is Ozgunes, Gandhi and Kassam, A variably trimmed mean CFAR radar detector, IEEE Transactions on Aerospace and Electronic Systems 28(4), 1002–1014. A later generalized smallest-of method based on trimmed means is described in A generalized smallest of selection CFAR algorithm.
Verification needs four layers: unit-test the sorting and retained indices; reproduce the worked-window estimates; calibrate false alarm under the assumed homogeneous law; and stress-test clutter edges, multiple targets, spillover and correlation. Report N, guard cells, L, H, retained count, multiplier, assumed clutter law and uncertainty in simulated performance.
CHAPTER 11 OF 12 · TRANSFER AND CONTINUE
11. Frequently asked questions
Why trim low samples as well as high samples?
Low trimming changes sensitivity to the low side of a heterogeneous window and can form part of a detector designed for clutter transitions. It can also bias the estimate if the low values are legitimate.
Is trimmed-mean CFAR cheaper than OS-CFAR?
Both require rank information in a direct implementation. Architecture, partial sorting and adaptive rules determine the real cost; do not infer it from the estimator name alone.
How many high cells should be removed?
Use a declared contamination and spillover model, then validate it. The trimming budget must count contaminated reference cells, not merely visible objects.
Can one multiplier work for every trim setting?
No. The statistic changes with L and H, so its calibration changes too.
CHAPTER 12 OF 12 · TRANSFER AND CONTINUE
12. Final transfer check and connected reading
Before approving a trimmed-mean detector, ask: Which cells are retained? What interference pattern is the trim intended to tolerate? Is spatial side information being lost by sorting? Was the multiplier calibrated for the exact statistic and dependence? Does the detector still meet false-alarm and detection requirements at edges and under spillover?
Continue through the Radar root, the order-statistic CFAR rank guide and CFAR loss and range-budget consequences. The transferable skill is to treat censoring depth, calibration and scene structure as one detector—not three independent settings.
Top · Contents · Previous · How X Works