---
title: 'HAROOD: Radar & Sensor HAR Benchmark'
url: https://www.emergentmind.com/topics/harood
type: topic
---

# HAROOD: Radar & Sensor HAR Benchmark

to=arxiv_search.search  微信的天天中彩票=json
{"query":"all:HAROOD", "max_results": 10, "sort_by":"submittedDate", "sort_order":"descending"}
to=arxiv_search.search  天天爱彩票是json
{"query":"ti:HAROOD OR abs:HAROOD", "max_results": 10, "sort_by":"relevance", "sort_order":"descending"}
to=arxiv_search.search ിക്കും to=json
{"query":"ti:HADRON OR abs:\"Halo mAss Distribution ReconstructiON\"", "max_results": 5, "sort_by":"relevance", "sort_order":"descending"}
to=arxiv_search.search 彩娱乐彩票 to=json
{"query":"ti:\"Haro 11\" OR abs:\"Haro 11\"", "max_results": 5, "sort_by":"relevance", "sort_order":"descending"}
HAROOD is a polysemous term in the arXiv literature. In its primary contemporary usage, it denotes two distinct research artifacts in human activity recognition: a short-range 60 GHz FMCW radar system for joint activity classification and out-of-distribution detection, and a benchmark for out-of-distribution generalization in sensor-based HAR. The same string also appears in contexts where it is likely a misspelling or shorthand for unrelated work, notably HADRON in large-scale-structure cosmology and Haro 11 in galaxy-merger studies. The meaning of HAROOD is therefore domain-specific rather than universal [2312.08894], [2512.10807].

## 1. Scope and principal meanings

The dominant technical meanings of HAROOD in recent arXiv usage are summarized below.

| Usage | Domain | Core characterization |
|---|---|---|
| HAROOD [2312.08894] | Radar HAR | Short-range FMCW radar-based human activity classifier and OOD detector |
| HAROOD [2512.10807] | Sensor-based HAR | Benchmark for OOD generalization in sensor-based human activity recognition |
| HADRON [1501.05520] | Cosmology | Halo mAss Distribution ReconstructiON |
| Haro 11 context [2503.01982] | Astrophysics | Separate usage associated with Haro 11 |

A common misconception is that HAROOD names a single model family. The literature instead uses the acronym for at least two unrelated HAR-oriented constructs. One is an end-to-end sensing and inference system built around macro and micro range-Doppler images, while the other is a modular evaluation framework spanning multiple datasets, domain shifts, algorithms, and model-selection protocols. This suggests that bibliographic interpretation of the term should begin with the sensing modality and experimental objective rather than with the acronym alone.

## 2. HAROOD as a short-range FMCW radar system

In the radar literature, HAROOD is a short-range FMCW radar-based human activity classifier and out-of-distribution detector designed for indoor environments such as homes, offices, and classrooms [2312.08894]. Its in-distribution label space consists of three human activities—sitting, standing, and walking—while out-of-distribution samples include any other moving or stationary object, including fans, vacuum cleaners, curtains, and empty-room clutter. The hardware platform is Infineon’s BGT60TR13C 60 GHz FMCW radar, with \(N_{Tx}=1\) and \(N_{Rx}=3\). The dataset uses humans at \(1\)–\(4\) m from the radar, a mounting height of \(2.5\) m, and a \(30^\circ\) downward tilt.

The system operates on two complementary radar representations: a macro RDI and a micro RDI. The macro RDI is formed from range FFT, mean removal, moving target indication, and Doppler FFT, and captures relatively coarse movement across range and Doppler. The micro RDI is formed by stacking eight range spectrograms, applying mean removal on both fast and slow times, applying Sinc filtering, and performing Doppler FFT along slow time; it emphasizes finer micro-Doppler signatures associated with limb and body-part motion. Both representations then undergo E-RESPD post-processing.

These design choices encode a dual-view representation of human motion. Standing concentrates energy around zero Doppler and fixed range bins, walking produces strong and varying Doppler components together with periodic limb signatures, and sitting yields short transition bursts with lower sustained micro-Doppler. A plausible implication is that HAROOD treats macro and micro RDIs as complementary observables of the same latent motion process, with the former emphasizing bulk kinematics and the latter emphasizing fine-grained articulatory structure.

## 3. Architecture, optimization, and OOD decision rule

HAROOD uses a two-stage architecture [2312.08894]. Stage 1 contains an encoder-decoder pair for the macro RDI, an encoder-decoder pair for the micro RDI, and a CNN plus linear block that consumes reconstructed macro and micro RDIs and outputs 1D embeddings. Stage 2 is a simple and lightweight convolutional classifier trained on the Stage-1 embedding with cross-entropy loss.

The defining aspect of Stage 1 is its joint objective:
$$
\mathcal{L}=\mathcal{L}_{\text{rec}}+\mathcal{L}_{\text{tri}}+\mathcal{L}_{\text{con}}.
$$
The reconstruction term is an intermediate reconstruction loss based on mean squared error. The triplet term structures the embedding so that anchor-positive pairs are closer than anchor-negative pairs, with margin \(\alpha_t=2\). The contrastive term functions as outlier exposure, using ID-OOD pairs and margin \(\alpha_c=2\), and is applied only to the encoders for the first 3 epochs of a 6-epoch Stage-1 schedule. Stage 1 uses Adamax; Stage 2 uses Adam.

OOD detection is based on reconstruction error rather than classifier confidence. The scalar OOD score is
$$
S=1\cdot \mathrm{MSE}_{\text{macro}}+0.001\cdot \mathrm{MSE}_{\text{micro}}.
$$
A threshold \(\tau\) is chosen so that about \(95\%\) of ID data have \(S \le \tau\); samples with \(S>\tau\) are flagged as OOD. The intended mechanism is explicit: encoder-decoder pairs are trained only on ID human activities, contrastive outlier exposure pushes OOD embeddings away from the ID manifold, and triplet loss compacts the ID manifold into better-separated sit, stand, and walk clusters. This suggests that HAROOD couples metric learning and reconstruction-based novelty detection rather than treating OOD scoring as a post hoc confidence calibration problem.

## 4. Dataset construction and empirical performance of the radar HAROOD

The radar HAROOD dataset contains 227,467 balanced ID training frames, an outlier-exposure set of 21,222 ID frames and 6,234 OOD frames from only two OOD types, and a balanced test set of 61,304 ID frames and 47,409 OOD frames [2312.08894]. The data were collected from 4 individuals across multiple indoor rooms, with 13 rooms used for training only and 8 rooms used for inference only. OOD test categories include table fan, stand fan, RC toy car, vacuum cleaner, robot vacuum, laundry, blinds, curtains, boiling water from a kettle, running water from a tap, and empty room with stationary clutter.

On activity classification, HAROOD reports an average accuracy of \(96.51\%\), with per-class accuracies of \(92.43\%\) for Sit, \(95.95\%\) for Stand, and \(99.83\%\) for Walk. The comparison ResNet-34 baseline reports \(94.38\%\) average accuracy, with \(91.38\%\), \(89.91\%\), and \(99.63\%\) for Sit, Stand, and Walk, respectively. The main gain is on Stand.

On OOD detection, HAROOD reports an average AUROC of \(95.04\%\). The per-activity AUROCs are \(93.57\%\) for Sit, \(95.43\%\) for Stand, and \(96.13\%\) for Walk. Additional metrics are reported as follows: for Sit, AUPR\(_{IN}=72.81\%\), AUPR\(_{OUT}=97.83\%\), and FPR95 \(=16.51\%\); for Stand, AUPR\(_{IN}=89.04\%\), AUPR\(_{OUT}=97.95\%\), and FPR95 \(=14.29\%\); for Walk, AUPR\(_{IN}=94.22\%\), AUPR\(_{OUT}=97.76\%\), and FPR95 \(=23.21\%\). Test time is \(28\) s for all test samples, compared with \(401\) s for ODIN, \(91\) s for MSP, \(91\) s for Energy, \(1480\) s for MAHA, \(1278\) s for FSSD, \(91\) s for OE, \(282\) s for GradNorm, \(97\) s for ReAct, and \(97\) s for MaxLogit. The paper states that HAROOD surpasses these baselines in almost all metrics.

The experimental narrative also identifies the main residual confusion: errors occur mostly between sit and stand, while walk is rarely misclassified due to its distinctive micro-Doppler. This is consistent with the stated representational role of macro and micro RDIs.

## 5. HAROOD as a benchmark for OOD generalization in sensor-based HAR

A separate arXiv work uses HAROOD to denote a benchmark and codebase for out-of-distribution generalization in sensor-based human activity recognition [2512.10807]. Its stated motivation is that realistic HAR deployments experience distributional shifts across individuals, devices, environments, and time, and that prior evaluations had covered only limited scenarios such as cross-device or cross-position. HAROOD addresses this by defining four OOD scenarios—cross-person, cross-position, cross-dataset, and cross-time—and constructing a testbed covering 6 datasets, 16 comparative methods, CNN-based and Transformer-based architectures, and two model-selection protocols.

The six datasets are DSADS, USC-HAD, UCI-HAR, PAMAP2, EMG, and WESAD. The benchmark formalizes the standard multi-source domain-generalization setup,
$$
\mathcal{D}^{tr}=\{\mathcal{D}^i\}_{i=1}^S,\qquad \mathcal{D}^i\sim P^i(\mathbf{x},y),
$$
with target-domain distribution \(P^T(\mathbf{x},y)\neq P^i(\mathbf{x},y)\) for all source domains. The goal is to learn a classifier using only source-domain data and minimize target error on the unseen domain.

The four scenarios are constructed as follows. Cross-person partitions subjects into disjoint domains and evaluates leave-one-domain-out transfer. Cross-position uses DSADS only, splitting each original \(45\times 1\times 125\) window into five \(9\times 1\times 125\) sub-windows corresponding to body locations. Cross-dataset harmonizes DSADS, USC-HAD, UCI-HAR, and PAMAP2 to common sensors, common positions, and six common activities, yielding windows of shape \(6\times 1\times 50\). Cross-time uses PAMAP2, EMG, and WESAD, applying sliding windows, sorting by recording time, and partitioning into four chronological segments treated as domains.

The 16 evaluated algorithms are ERM, Mixup, DDLearn, DANN, CORAL, MMD, VREx, LAG, MLDG, RSC, GroupDRO, ANDMask, Fish, Fishr, URM, and ERM++. HAROOD implements each with two backbones, a CNN and a Transformer, within a modular PyTorch framework. It also standardizes model selection through a training-domain validation protocol and an oracle protocol. The benchmark reports that 88,320 neural networks are trained.

## 6. Benchmark findings, model-selection effects, and methodological significance

The principal empirical finding of the HAROOD benchmark is that no single method consistently outperforms others across all datasets, scenarios, backbones, and model-selection protocols [2512.10807]. This directly counters the idea that OOD generalization in HAR is solved by one preferred algorithmic family. On DSADS cross-person with CNN, CORAL is often best; on UCI-HAR cross-person, URM is top under some settings; on DSADS cross-position, plain ERM is surprisingly strong; on WESAD cross-time, ERM++ with CNN performs extremely well. Under Transformer backbones, the ranking changes materially: LAG and ANDMask move into top positions, while ERM++ drops sharply.

Another central conclusion is that OOD-specific algorithms do not uniformly beat ERM. Methods such as MLDG and DANN often fail to surpass ERM and sometimes perform much worse, while strong baselines such as ERM++ can rival or beat many specialized OOD methods. By contrast, CORAL, Fish, and Fishr emerge as robust choices across many settings. The benchmark therefore reframes the problem from one of universal algorithmic dominance to one of scenario- and architecture-specific trade-offs.

Model selection is itself a substantive variable. Oracle selection improves average performance by about \(+2\) percentage points, but does not radically change ranking order. This suggests that part of the difficulty in OOD HAR lies in hyperparameter sensitivity and in the mismatch between training-domain validation and truly OOD deployment. The benchmark further supports its conclusions with rank-sum aggregation, Wilcoxon signed-rank tests, confusion matrices, and t-SNE visualizations. A plausible implication is that future progress may depend at least as much on OOD-aware selection protocols and architecture-method coupling as on new loss functions alone.

## 7. Related systems and terminological ambiguity

Within radar HAR more broadly, HAROOD sits alongside adjacent infrastructure rather than exhausting the field. RadHARSimulator V1, for example, is a model-based FMCW radar HAR simulator that integrates a \(13\)-scatterer anthropometric kinematic model, simulates \(12\) distinct activities, supports free-space and through-the-wall propagation, and produces RTMs and DTMs via STFT and FSST [2509.06751]. Its role is different from HAROOD’s: it is a synthetic data generator and validation platform rather than a benchmark for OOD generalization or a short-range classifier/OOD detector for three human activities. This contrast helps delimit what HAROOD means in radar work: either a specific radar inference architecture or a broader OOD benchmark, not a general label for all radar HAR tooling.

The acronym is also a source of cross-domain ambiguity. In cosmology, a similar string may refer to HADRON, the Halo mAss Distribution ReconstructiON method for assigning halo masses in large-scale-structure mocks [1501.05520]. In astrophysics, the string has also been used in explanatory context as shorthand associated with Haro 11, the local starburst galaxy modeled through hydrodynamical merger simulations [2503.01982]. These are terminologically adjacent but scientifically unrelated usages.

Taken together, the literature supports a narrow encyclopedia definition: HAROOD refers chiefly to either a two-stage radar-based system for simultaneous human activity classification and OOD detection or a comprehensive benchmark for OOD generalization in sensor-based HAR. The shared acronym masks substantial differences in sensing modality, objective function, data regime, and evaluation philosophy.

Source: https://www.emergentmind.com/topics/harood