Low-cost low-power in-vehicle occupant detection with mm-wave FMCW radar

George Shaker; Hajar Abedi; Mostafa Alizadeh

REVIEW 4 major objections 6 minor 7 references

Low-cost low-power in-vehicle occupant detection with mm-wave FMCW radar

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper reports that a low-cost 77 GHz FMCW radar, using Capon range-azimuth maps and an SVM classifier, detects in-vehicle occupancy by row with 97.8% average accuracy and 100% accuracy on the empty-versus-occupied decision.

desk verdict Useful application note, but the empirical claims are not yet established because the evaluation protocol is under-reported and possibly leaky. read the letter →

arxiv 1908.04417 v1 pith:UDPBZ6RP submitted 2019-08-12 eess.SP eess.IV

Mostafa Alizadeh , Hajar Abedi , George Shaker This is my paper · ORCID

classification eess.SPeess.IV

keywords FMCWradarin-vehicleoccupantdetectionCaponfilterrange-azimuthestimationsupportvectormachinemm-waveTDMMIMOvehicleoccupancy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a cheap, low-power millimeter-wave FMCW radar can do reliable in-vehicle occupant detection without the cost or power of high-resolution sensors. It proposes a signal-processing chain: range FFT, Capon filtering for joint range-azimuth estimation, PCA to shrink the feature maps, and a support vector machine to label eight occupancy scenarios in a minivan. On its own experiments the system reports 97.8% average accuracy over the eight classes and 100% accuracy on the binary question of whether the car is empty. The practical target is an IoT-friendly sensor that knows which row of a vehicle is occupied, useful for child-presence alerts, safety systems, and smart-cabin features.

What carries the argument

The central object is the range-azimuth map produced by the Capon filter, a minimum-variance beamformer that estimates the angle-of-arrival spectrum while suppressing interference. The map is built per range bin from the covariance matrix $R$ of the eight virtual receiver channels using $\Phi(\hat{\theta}) = 1/[a^H(\hat{\theta}) R^{-1} a(\hat{\theta})]$, where $a(\hat{\theta})$ is the steering vector for a test angle. Because the steering matrix in the covariance model is Vandermonde, $R$ is positive definite and therefore invertible, so the Capon output is well defined. The resulting range-azimuth maps replace the one-dimensional range profile as the input to PCA and then to the SVM, giving the classifier spatial structure to work with despite poor angular resolution.

What would settle it

Collect a new test set in a different minivan with different subjects, retrain with the same 8-to-2 protocol, and check whether empty-versus-occupied accuracy stays at 100% and average row accuracy stays near 97.8%; a large drop would show the reported accuracy is an artifact of the original recording setup. A cheaper check is to inspect whether PCA was fit only on the training portion of the split; if it used the full dataset, the reported test accuracy is optimistically biased.

Watch

Extended reading notes

Core claim

The central discovery is that a 77 GHz FMCW radar with only 14 degrees of azimuth resolution (about 25 cm at one meter) can still localize occupants to rows when raw range profiles are converted into range-azimuth maps by a Capon beamformer and those maps are fed to an SVM. The authors define eight classes—empty car, one person in row 1, 2, or 3, one person in each of two rows, and one person in each of three rows—and their classifier separates them, with no false alarms when the car is empty and no occupied case reported as empty. The authors attribute the success to using joint range-azimuth information rather than a one-dimensional range profile, plus PCA to keep the feature set minimal. Their claim is that intelligent feature extraction can make low-resolution, low-cost radar sufficient for row-level occupancy detection.

Load-bearing premise

The accuracy figures stand or fall on the assumption that the evaluation data are independent and representative: if the 8-to-2 split or the 5-fold cross-validation leaks information—for example, the same person appearing in both training and test sets, or PCA fit on the full dataset—the 97.8% and 100% numbers will not reproduce for new people, vehicles, or sessions.

Editorial extensions

If this is right

A single low-cost automotive radar chip can supply presence and row-level occupancy information without cameras, making child-presence alerts and seat-belt reminders feasible in any trim level.
The 100% accuracy on the empty-versus-occupied decision means a binary “someone is still in the car” alert is reliable under the tested conditions.
Reducing the problem from all 128 possible seat combinations to eight row-level classes keeps the classifier tractable while still covering the realistic single-passenger and multi-passenger cases.
After PCA, the feature dimension is small enough that the SVM step is computationally light, which matters for embedded vehicle hardware.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension is to compare the Capon range-azimuth maps against ordinary FFT-based beamforming maps with the same SVM; that would show how much of the accuracy gain comes from the Capon filter itself.
The pipeline likely transfers to other confined cabins such as SUVs or truck cabs, but the 97.8% figure is tied to this minivan's geometry, seat layout, and recording conditions, so per-vehicle retraining would be needed.
The error pattern visible in the confusion matrix—localization of three occupants is the hardest case—suggests a hierarchical design that first detects empty versus occupied and then localizes occupants might be more robust than a single eight-way classifier.
A stricter evaluation that splits by subject or by recording session, rather than a random 8-to-2 split, would test whether the method generalizes across people and days.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The core idea is sensible: take a low-cost 77 GHz FMCW radar, form Capon range-azimuth maps, reduce with PCA, and let an SVM sort the cabin into eight occupancy scenarios. That is a legitimate engineering contribution, and the paper's main observation—that the SVM can compensate for the coarse 14-degree azimuth resolution—is worth recording. The Capon derivation in Section II is standard and correctly executed; there is no circularity in the signal model itself.

What the paper does well is the system integration and the choice of a realistic test environment (a minivan with occupants moving, talking, using phones). The confusion matrix suggests the empty-versus-occupied detection is genuinely robust, which matters for child-presence and airbag applications.

Now the soft spots, and they are load-bearing because the entire empirical contribution rests on two numbers: 97.8% scenario accuracy and 100% empty-versus-occupied accuracy. First, the order of operations in Section II-B is ambiguous at best: PCA is described as reducing the dataset before the 8:2 split. If PCA is fit on the full dataset, the retained components have seen the test set, which inflates any subsequent cross-validation accuracy. The paper never states explicitly that PCA was fit only on the training folds. Second, the relationship between the 8:2 split and the 5-fold cross-validation is unclear. The text says the split is used to create training and test sets, then says evaluation used 5-fold CV. Which number is reported—test-set accuracy or CV accuracy? Third, there is no dataset size, no number of subjects or recording sessions, and no per-fold or per-class variance. Those omissions make it impossible to judge whether 97.8% would survive another vehicle or another day.

These are fixable, but they are not minor. The authors need to say how PCA was embedded in the cross-validation, report the actual number of recordings per class, and give fold-level results. I would also like to see a baseline comparison—for instance, classification from range-only features or from raw range-azimuth maps without PCA—to justify that the Capon azimuth information is what helps.

The math and the idea are sound, and the paper is a serious candidate for peer review if the evaluation is tightened. As it stands, I would not cite the accuracy numbers yet, but I would send it to a careful referee who can push for the missing details. The right outcome is a conditional accept with mandatory revisions on the evaluation protocol.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes an in-vehicle occupant detection system using a 77 GHz FMCW radar. The signal processing chain is: range FFT on the received chirps, stationary clutter removal, covariance-matrix estimation across the virtual channels per range bin, 2D Capon range-azimuth spectrum, PCA dimensionality reduction, and SVM classification into eight occupancy scenarios (empty car, each row occupied, and multi-row combinations). Experiments were carried out in a minivan with a TI radar (3Tx/4Rx, two transmitters in TDM mode giving eight virtual receivers). The paper reports 97.8% average accuracy with 5-fold cross-validation and 100% accuracy for distinguishing empty versus occupied. The Capon derivation is standard, but the empirical reporting is incomplete in ways that affect the validity of the headline claims.

Significance. If the reported accuracy is reproducible, the work is practically valuable: it suggests that a low-cost, low-power mm-wave radar combined with a Capon/SVM pipeline can detect occupant presence and coarse row localization in a vehicle, with potential applications in automotive safety, child-presence detection, and IoT sensing. The Capon range-azimuth formulation is standard and is not circular; no fitted quantity is fed back into the accuracy claim. The use of eight clearly defined occupancy classes is a useful framing. However, the empirical contribution is not yet fully validated because the evaluation protocol is under-specified and may suffer from feature leakage. The practical significance therefore depends on the outcome of a properly described and executed cross-validation procedure.

major comments (4)

[II-B, III] Section II-B states that PCA is applied to reduce the size of the range-azimuth dataset before the 80/20 split, and Section III then reports evaluation with 5-fold cross-validation. If PCA is fit on the entire dataset before any split, the test folds contribute to the feature extractor, violating the independence of training and test data and potentially inflating the reported 97.8% and 100% accuracies. This is load-bearing because these accuracies are the entire empirical contribution. Please specify whether PCA was fit only on the training portion within each cross-validation fold; if it was not, re-run the evaluation with PCA nested inside the cross-validation loop and report the resulting figures.
[III, experimental setup] The manuscript does not report the number of recordings, subjects, sessions, or samples per class. The statement that recording duration is the same for all classes does not quantify the dataset. Without these numbers, the reader cannot assess class balance, the statistical uncertainty of the 97.8% figure, or whether the test set is representative of the intended operating conditions (e.g., different occupants, seat positions, movements, or sessions). Please provide the total dataset size, the number of samples per class, the number of distinct participants, and the number of recording sessions.
[III, accuracy reporting] The text is ambiguous about whether 97.8% comes from the 80/20 split described in Section II-B or from the 5-fold cross-validation: Section II-B describes a split 'by the ratio of 8 to 2,' while Section III concludes 'by using 5-fold cross validation, 97.8% correct detection rate is obtained.' Please clarify which protocol produced the reported number, report the per-fold accuracies, and specify whether the confusion matrix in Fig. 4 is aggregated over all test folds or corresponds to a single holdout split.
[II-B, hyperparameter selection] The SVM hyperparameters are selected by grid search with k-fold cross-validation, and the model is then evaluated with 5-fold cross-validation. Please clarify whether the hyperparameter search was conducted inside each training fold of the 5-fold evaluation (nested cross-validation) or on the full dataset. If the latter, the reported accuracy may be optimistically biased because the test folds would have influenced model selection.

minor comments (6)

[II-A] There is an incomplete sentence at the end of Section II-A (the fragment 'In an') that should be removed or completed.
[II-B] The phrase 'principle component analysis' should be 'principal component analysis'; the same typo appears in the Introduction.
[Fig. 4] The confusion matrix in Fig. 4 has no axis labels; please label the axes with the class names from Table I so that the claims about the first column are directly verifiable.
[Table II] The row/column headers of Table II are garbled; they should be readable as 'chirp slope, chirp duration, sweeping bandwidth, frame rate, ADC sampling rate' with their corresponding values.
[Eqs. (1)-(4)] The notation is not fully defined: the signal s(ts) in Eq. (4) is not defined earlier, and the function y(v, fb, tf, ts) in Eq. (2) has an inconsistent argument list relative to its use in Eq. (3). Please define all symbols consistently.
[Figs. 2 and 3] Figure 2's caption 'Minivan indoor look' is unclear, and Fig. 3 would benefit from labeled axes (range and azimuth) and a colorbar to make the range-azimuth map interpretable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: measured accuracy and external Capon formulation make the central claim independent of its inputs.

full rationale

The paper's central claim is an empirical accuracy result obtained by training an SVM on range-azimuth features and testing it with cross-validation. The Capon spectrum in Eq. (5), Phi(theta-hat) = 1 / (a^H(theta-hat) R^{-1} a(theta-hat)), is a standard external signal-processing estimator; it is not derived from the occupancy labels or from the reported accuracy. The 97.8% and 100% figures are presented as measured outcomes, not as quantities that are fitted then re-predicted. The PCA step is described as dimensionality reduction 'without information loss' before the train/test split, and while the evaluation-protocol details are ambiguous, that is a data-leakage or reproducibility concern, not a circularity of the kind where an input is equivalent to the output by construction. The self-citations in the introduction ([1], [4]) are background references for mm-wave radar applications and are not load-bearing for the claimed detection result. No equation is defined in terms of the target result, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. Therefore the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The core physics is standard FMCW beamforming; the only fitted quantity is the SVM hyperparameter set. The main experimental assumptions concern the validity of the clutter model and, more importantly, that the single-vehicle recordings support generalization. No new physical entities are introduced.

free parameters (1)

SVM hyperparameters = not reported
A grid search with k-fold cross-validation was used to select classifier hyperparameters (Section II-B). The chosen values are not listed, so the reported accuracy depends on settings that are not independently reproducible.

assumptions (5)

domain assumption FMCW point-target signal model of Eq. (1)
The Capon range-azimuth derivation assumes the received baseband signal comes from a point target with the given phase terms and additive noise. Real extended human bodies and multipath deviate from this model.
domain assumption Independent receiver noise with positive-definite covariance Rn
Used to justify invertibility of R in Eq. (4); this is a standard mild assumption.
standard math A(theta) is a Vandermonde matrix with positive kernels, making the signal covariance positive definite
Invoked in Section II-A with citation [7] to guarantee R is invertible; this requires distinct angles and an adequate number of antenna elements.
domain assumption Stationary clutter removal by subtracting the average of each range bin
Section II assumes that subtracting the mean eliminates stationary scatterers so the remaining signal is dominated by moving occupants.
domain assumption The recording conditions in one minivan represent deployment conditions
Section III lets passengers move, talk, and use phones, but uses one vehicle and an unreported number of subjects and sessions; generalization to other vehicles is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Low-cost low-power in-vehicle occupant detection with mm-wave FMCW radar." pith.science (2026). https://pith.science/paper/UDPBZ6RP

@misc{pith2026190804417,
  author       = {Pith},
  title        = {Pith review of: Low-cost low-power in-vehicle occupant detection with mm-wave FMCW radar},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UDPBZ6RP}},
  note         = {Machine review of arXiv:1908.04417}
}

read the original abstract

In this paper, we use a low-cost low-power mm-wave frequency modulated continuous wave (FMCW) radar for the in-vehicle occupant detection. We propose an algorithm using Capon filter for the joint range-azimuth estimation. Then, the minimum necessary features are extracted to train machine learning classifiers to have reasonable computational complexity while achieving high accuracy. In addition, experiments were carried out in a minivan to detect occupancy of each row using support vector machine (SVM). Finally, our proposed system achieved 97.8% accuracy on average in finding the defined scenarios. Moreover, the system can correctly identify if the vehicle is occupied or not with 100% accuracy.

Figures

Figures reproduced from arXiv: 1908.04417 by the authors.

**Figure 1.** Vehicle occupant detection algorithm. where tf and ts are referring to two different time scales. The former refers to the time scale of a chirp period while the latter is referring to the time scale greater than a chirp period. bl , fb, v, λmax, τl , ξl , ∆ψl(tf , ts), and el(tf , ts) are the channel gain, beat frequency 1 , target’s velocity, the wavelength corresponding to the start frequency of the FMCW ramp, th… view at source ↗

**Figure 2.** Minivan indoor look work with their phones. We consider eight classes, empty car and either a row is occupied or not (Table I). In fact, our class definition is a small collection of all possible situations in a car; otherwise, the total number of seat occupations is 2 7 = 128. The recording duration is the same for all classes in order to have unbiased training. TI radar chip has 3Tx and 4Rx and two transmitters we… view at source ↗

**Figure 3.** Range-azimuth map when the seat number 1, 4, and 5 are occupied. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 7 canonical work pages

[1]

Remote heart rate sensing with mm-wave radar,

M. Alizadeh, G. Shaker, and S. Safavi-Naeini, “Remote heart rate sensing with mm-wave radar,” in 2018 18th International Symposium on Antenna Technology and Applied Electromagnetics (ANTEM) , pp. 1–2

work page 2018
[2]

Wireless biometric individual identiﬁcation utilizing millimeter waves,

K. Diederichs, A. Qiu, and G. Shaker, “Wireless biometric individual identiﬁcation utilizing millimeter waves,” vol. 1, no. 1, pp. 1–4

work page
[3]

Gesture recognition using mm-wave sensor for human-car interface,

K. A. Smith, C. Csech, D. Murdoch, and G. Shaker, “Gesture recognition using mm-wave sensor for human-car interface,” IEEE Sensors Letters , vol. 2, no. 2, pp. 1–4, June 2018

work page 2018
[4]

Remote monitoring of human vital signs using mm-wave fmcw radar,

M. Alizadeh, G. Shaker, J. C. M. D. Almeida, P. P. Morita, and S. Safavi- Naeini, “Remote monitoring of human vital signs using mm-wave fmcw radar,” IEEE Access, vol. 7, pp. 54 958–54 968, 2019

work page 2019
[5]

Principal component analysis-based occupancy detection with ultra wideband radar,

Z. Baird, I. Gunasekara, M. Bolic, and S. Rajan, “Principal component analysis-based occupancy detection with ultra wideband radar,” in 2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS), Aug 2017, pp. 1573–1576

work page 2017
[6]

Short-range millimetric-wave radar system for occupancy sensing application,

A. Santra, R. V . Ulaganathan, and T. Finke, “Short-range millimetric-wave radar system for occupancy sensing application,” IEEE Sensors Letters , vol. 2, no. 3, pp. 1–4, Sep. 2018

work page 2018
[7]

Spectral properties of totally positive kernels and matrices,

A. Pinkus, “Spectral properties of totally positive kernels and matrices,” in Total positivity and its applications . Springer, 1996, pp. 477–511

work page 1996

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Remote heart rate sensing with mm-wave radar,

M. Alizadeh, G. Shaker, and S. Safavi-Naeini, “Remote heart rate sensing with mm-wave radar,” in 2018 18th International Symposium on Antenna Technology and Applied Electromagnetics (ANTEM) , pp. 1–2

work page 2018

[2] [2]

Wireless biometric individual identiﬁcation utilizing millimeter waves,

K. Diederichs, A. Qiu, and G. Shaker, “Wireless biometric individual identiﬁcation utilizing millimeter waves,” vol. 1, no. 1, pp. 1–4

work page

[3] [3]

Gesture recognition using mm-wave sensor for human-car interface,

K. A. Smith, C. Csech, D. Murdoch, and G. Shaker, “Gesture recognition using mm-wave sensor for human-car interface,” IEEE Sensors Letters , vol. 2, no. 2, pp. 1–4, June 2018

work page 2018

[4] [4]

Remote monitoring of human vital signs using mm-wave fmcw radar,

M. Alizadeh, G. Shaker, J. C. M. D. Almeida, P. P. Morita, and S. Safavi- Naeini, “Remote monitoring of human vital signs using mm-wave fmcw radar,” IEEE Access, vol. 7, pp. 54 958–54 968, 2019

work page 2019

[5] [5]

Principal component analysis-based occupancy detection with ultra wideband radar,

Z. Baird, I. Gunasekara, M. Bolic, and S. Rajan, “Principal component analysis-based occupancy detection with ultra wideband radar,” in 2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS), Aug 2017, pp. 1573–1576

work page 2017

[6] [6]

Short-range millimetric-wave radar system for occupancy sensing application,

A. Santra, R. V . Ulaganathan, and T. Finke, “Short-range millimetric-wave radar system for occupancy sensing application,” IEEE Sensors Letters , vol. 2, no. 3, pp. 1–4, Sep. 2018

work page 2018

[7] [7]

Spectral properties of totally positive kernels and matrices,

A. Pinkus, “Spectral properties of totally positive kernels and matrices,” in Total positivity and its applications . Springer, 1996, pp. 477–511

work page 1996