NUP-REPORT 1.0: A Minimum Reporting and Benchmarking Framework for Non-Upright Pedestrian Detection and Pre-Crash Safety Evaluation

1. Introduction

Road traffic injury remains a major global public-health problem, and pedestrians continue to experience disproportionate risk [1]. Within this population, people lying on the roadway constitute a small but exceptionally severe subgroup. Japanese nationwide data showed that pedestrians lying on the road accounted for 8.3% of pedestrian fatalities but only 0.6% of all pedestrian casualties during 2012–2016 [2]. A subsequent analysis of 2,452 such collisions reported that more than 80% occurred at night and identified impact speed, body-region injury, vehicle type and hit-and-run involvement as important correlates of fatal or severe outcome [3]. These findings establish a clear prevention need but do not, by themselves, specify how vehicle perception systems should be evaluated.

Pedestrian detection has advanced through large datasets, stronger convolutional and transformer architectures, multispectral sensing and multimodal fusion. Canonical benchmarks have improved evaluation of scale, occlusion and urban diversity [4,5,6,7,8,9,10,11]. Behavioural datasets have also enabled modelling of crossing intent and pedestrian action [12,13]. However, the dominant pedestrian representation in these resources remains an upright or approximately upright person. A road-level human body may occupy a wider and lower image region, present weaker conventional shape cues, overlap with debris or shadows, and exhibit a thermal or radar signature that differs from the standing class. The resulting problem is not only lower detection performance. It is also uncertain comparability: two studies may report similar recall while testing different postures, ranges, target surrogates, illumination, sensor states, data partitions and latency boundaries.

Recent work has begun to address fallen-person detection and its safety implications. A multimodal architecture combining long-wave infrared, near-infrared stereo and ultrasonic sensing was evaluated for falling and fallen-person detection [14]. A physics-grounded multimodal framework subsequently addressed kinematic reconstruction and uncertainty propagation after pedestrian impact [15]. Simulation-based research connected non-upright detection to braking and exploratory injury-risk estimation [16], while an integrative preprint linked detection, intervention and forensic reconstruction within a broader prevention architecture [17]. Collectively, these studies suggest that perception, timing, actuation and physical consequences should be interpreted as a chain rather than as isolated results. They do not yet provide a general reporting framework that other research groups can apply independently of a particular architecture.

The need for explicit reporting is reinforced by safety-critical machine-learning research. Modern neural networks can be poorly calibrated [18], predictive uncertainty may degrade substantially under distribution shift [19], and out-of-distribution inputs can remain confidently misclassified [20]. Multimodal systems are also vulnerable to asymmetric sensor degradation, especially in adverse weather [21]. Safety assurance, therefore, requires more than reporting mean average precision or a single overall accuracy. It requires evidence about data provenance, condition-specific failure, calibration, latency and the connection between perception output and vehicle response [22,23,24,25].

This article introduces NUP-REPORT 1.0, a minimum reporting and benchmarking framework for non-upright pedestrian detection and pre-crash safety evaluation. Its objectives are to:

1. Define an operational vocabulary for non-upright targets;

2. specify the minimum information needed to interpret and reproduce a study;

3. distinguish perception accuracy from safety-relevant timing and physical response;

4. propose a minimum benchmark scenario matrix and core outcome set;

5. map the framework to current safety standards and assessment programmes without claiming equivalence or compliance; and

6. Provide a versioned checklist that can be refined through future empirical and multidisciplinary work.

Figure 1 summarises the six-domain NUP-REPORT reporting chain, from target definition to vehicle-level safety interpretation.

2. Scope, Terminology and Framework Development

2.1. Intended Use

NUP-REPORT is intended for research evaluating systems that detect a person in a non-upright configuration in or near a vehicle path. It may be applied to visible-spectrum cameras, thermal imaging, radar, LiDAR, ultrasonic sensing, event cameras, multimodal fusion or other relevant modalities. It is suitable for controlled-track experiments, laboratory studies, naturalistic datasets, simulation, digital twins and hybrid validation programmes.

The framework is designed for a mixed audience comprising computer-vision researchers, automotive engineers, injury-prevention and forensic researchers, test laboratories, consumer-assessment organisations, regulators and standards specialists. It can be used prospectively when designing a study or retrospectively when appraising an existing publication.

NUP-REPORT is not:

• a certification or homologation standard;
• a replacement for Euro NCAP, UNECE, NHTSA or ISO procedures;
• a validated performance threshold;
• evidence that a particular detector is deployment-ready;
• a requirement to test every conceivable scenario; or
• a claim that modelled stopping or injury outcomes represent observed real-world benefit.

2.2. Operational Definition of a Non-Upright Pedestrian

A non-upright pedestrian is defined here as a person whose observable body configuration differs materially from the standing, normally walking or running posture assumed in conventional pedestrian-detection datasets and vehicle test targets. The term describes posture, not cause. It should therefore not be replaced by causal labels such as “intoxicated”, “unconscious” or “medically incapacitated” unless those states are independently established.

Table 1 provides the minimum recommended posture vocabulary. Studies may add subclasses, but they should preserve the observable configuration and orientation information needed for comparison.

2.3. Framework-Development Method

NUP-REPORT was developed as a structured narrative framework, not a systematic review or formal consensus guideline. Source material was selected from five evidence streams through July 2026:

1. epidemiology and injury severity of pedestrians lying on the roadway [2,3];

2. canonical pedestrian, traffic-scene and multispectral benchmarks [4,5,6,7,8,9,10,11,12,13,21,26];

3. calibration, uncertainty and out-of-distribution research [18,19,20,22,23,27];

4. scenario-based development, testing and assurance of automated vehicles [24,25,28,29]; and

5. public standards, regulations and assessment protocols relevant to functional safety, SOTIF, scenarios, pedestrian targets and AEB [30,31,32,33,34,35,36,37,38,39].

A failure-chain decomposition was applied from target representation → sensing and data → model inference → uncertainty and timing → vehicle response → safety interpretation. Candidate reporting items were retained when omission could plausibly:

• prevent independent reproduction;
• invalidate cross-study comparison;
• conceal a safety-critical subgroup failure;
• make a latency claim ambiguous;
• confound measured and modelled evidence; or
• overstate the relationship between perception accuracy and collision avoidance.

The six resulting domains were then mapped against current standards and public assessment materials to clarify complementarity and boundaries. No Delphi panel, formal stakeholder vote or multi-laboratory validation was undertaken. Consequently, Version 1.0 is presented as a testable proposal for community refinement.

3. Why Conventional Reporting Is Insufficient

3.1. Dataset Representation and Hidden Pooling

Pedestrian datasets differ in geographic coverage, camera placement, target scale, occlusion, annotation policy and train/test construction [4,5,6,7,8,9,10,11]. Those differences already complicate comparison for upright detection. Non-upright detection introduces additional variation in torso orientation, projected aspect ratio, limb configuration, contact with the road plane and resemblance to non-human objects.

A study can appear large while containing limited independent evidence. Frame-level random splitting may place adjacent frames from the same sequence, the same actor, the same location or the same rendered scene in both training and testing. This can inflate apparent generalisation. NUP-REPORT therefore treats the sequence, event, actor, location and synthetic seed as relevant units of independence, not merely the image.

Synthetic data are valuable because rare and hazardous scenarios are difficult to collect. However, synthetic and real samples should not be pooled without disclosure. Rendering engine, sensor model, target model, domain randomisation, noise assumptions and the role of synthetic data in training and testing must be reported. A result obtained entirely from simulation should not be presented as real-world validation.

3.2. Pooled Accuracy Can Conceal the Safety-Critical Subgroup

Overall accuracy, precision, recall and mean average precision remain useful, but they can conceal poor performance in rare subgroups. For example, a dataset dominated by upright daytime targets may yield high aggregate recall even if prone nighttime targets are frequently missed. A pooled average also hides whether performance deteriorates under rain, reflective roads, occlusion, sensor contamination or modality loss.

NUP-REPORT therefore requires stratification by the conditions actually tested. At a minimum, posture-specific sample counts and outcomes should be provided. When sample size permits, posture should be crossed with range, illumination, occlusion and sensor state. Untested cells should be visible rather than implicitly treated as covered.

3.3. Confidence Is Not Equivalent to Reliability

A detector confidence score is not automatically a calibrated probability of correctness. Calibration is especially important when confidence controls sensor weighting, temporal confirmation, warning or braking. Expected calibration error, reliability diagrams, Brier score or other appropriate measures should therefore be reported when confidence is safety-relevant [18]. Calibration should be evaluated separately under nominal and shifted conditions because apparently well-calibrated models can degrade under distribution shift [19].

Out-of-distribution detection, selective prediction or abstention mechanisms may be useful, but their own false-negative and latency consequences must be reported. A system that abstains safely in a research benchmark may still leave the vehicle without sufficient time to intervene.

3.4. Inference Time Is Not End-to-End Latency

Studies often report neural-network inference time while omitting sensor exposure, readout, pre-processing, synchronisation, temporal confirmation, fusion, communication and actuator delay. For a moving vehicle, each component consumes distance. A 20 ms inference time does not imply a 20 ms perception-to-brake response.

NUP-REPORT therefore distinguishes:

• time to first detection (TTFD): delay from the first eligible sensor observation to the first valid detection;
• inference latency: model execution time for a defined hardware and input configuration;
• decision latency: additional temporal confirmation, tracking and decision logic;
• actuation latency: command transfer, brake-system delay and force build-up; and
• total response delay: the combined delay relevant to stopping-distance analysis.

Figure 2 illustrates the required separation of detection availability, inference and decision processing, system-response delay and physical deceleration.

4. The Six NUP-REPORT Domains

Table 2 summarises the six domains and identifies minimum and extended items. “Minimum” indicates information normally required to interpret a study. “Extended” indicates information strongly recommended when the corresponding method or claim is present.

4.1. Domain 1: Target and Posture

Every study should state exactly what counts as a positive target. Bounding-box aspect ratio alone is insufficient because a prone person, a bag, a blanket, road debris and a shadow may have similar two-dimensional geometry. The negative dataset should therefore include hard negatives that resemble a road-level human signature in one or more modalities.

Human participants, mannequins, articulated pedestrian targets, digital humans and simplified synthetic objects should be reported separately. When a surrogate is used, the study should describe which properties are human-representative and which are not. A thermally unrepresentative mannequin, for example, cannot establish thermal-detection performance without additional justification.

For falls or collapses, frame-level labels should distinguish the transition from the final stable posture. A prediction claim should identify the event definition and lead time relative to a clearly stated transition point.

4.2. Domain 2: Scenario and Environment

Scenario variables determine both detectability and physical intervention opportunity. Minimum reporting should include vehicle speed, target range, lane position, target orientation, movement state, illumination, weather, road surface, occlusion and background clutter.

Illumination should be quantified when possible rather than described only as “day” or “night”. The weather should distinguish the presence and severity of rain, fog or snow. Wet and reflective road surfaces should be reported because they may alter visible, near-infrared, thermal and LiDAR response. Headlamp state, direct glare and sensor contamination should be reported when relevant.

The initial detection distance should be measured relative to a defined target reference point. When lateral offset or road curvature materially affects the line of sight, a simple longitudinal distance may be insufficient, and the coordinate convention should be stated.

4.3. Domain 3: Sensors and Data Provenance

Sensor reporting should permit another laboratory to understand the physical measurement process. At minimum, studies should state modality, resolution or angular resolution, field of view, frame or scan rate, mounting position, calibration method and synchronisation procedure.

Data provenance should identify the source, collection period, geographic context, annotation procedure and unit of independence. Real, augmented and synthetic data should be separately enumerated. If public datasets are combined, the contribution of each source to training, validation and testing should be stated.

Sequence-level leakage is a major concern. Adjacent frames, repeated actors, near-identical simulations or the same location under minor variation should not be split randomly across training and testing without a sensitivity analysis. The preferred approach is group-wise partitioning by event, actor, location and synthetic seed.

4.4. Domain 4: Model and Fusion Architecture

The study should report model family and version, input resolution, pretraining source, fine-tuning procedure, class definitions, confidence threshold, non-maximum-suppression settings, tracking method, temporal confirmation, and compute hardware.

Multimodal research should include modality-specific baselines. The fused model should not be compared only with a weak or differently trained single-camera baseline. Ablation should remove or degrade one modality at a time while holding other factors constant. Missing-modality handling should also be disclosed: zero-filling, learned masking, fallback model, confidence reweighting and system shutdown are not equivalent strategies.

Latency should be reported for the deployed numerical precision and hardware. Batch-processing throughput is not equivalent to single-sample real-time latency. Warm-up, data transfer and pre/post-processing should be included or explicitly excluded.

4.5. Domain 5: Performance and Uncertainty

At minimum, precision, recall, false-negative count, sample size and uncertainty interval should be provided for each principal posture class. When the task is object detection, the intersection-over-union threshold and matching policy should be specified. For sequence-based detection, the valid-detection persistence criterion should be stated.

A safety-critical false negative is not simply any missed frame. It is a target that remains undetected beyond a pre-defined intervention boundary. Studies should therefore report both frame-level metrics and event-level safety-critical miss rates.

Calibration should be evaluated when confidence affects fusion or intervention. Reliability diagrams and ECE are useful, but binning choices and sample counts should be reported. Calibration should be stratified by posture or condition where practical.

Failure analysis should show representative false negatives and false positives, including modality-specific evidence. Failure cases should not be limited to visually obvious examples. Near-threshold and high-confidence failures are particularly informative.

4.6. Domain 6: Vehicle-Level Safety Interpretation

A detection result becomes safety-relevant only through time, distance and vehicle response. Studies that estimate collision avoidance or mitigation should report initial vehicle speed, detection distance, TTFD, total delay, effective deceleration, friction/gradient assumptions, brake-force build-up and whether steering is considered.

Measured vehicle tests, simulation and analytical calculations must be labelled separately. Modelled residual impact speed or injury probability should not be described as observed effectiveness. Sensitivity analysis should vary the assumptions that dominate the result, especially detection distance, delay and deceleration.

5. Minimum Benchmark Scenario Matrix

The minimum matrix is intended to reveal coverage, not to require a fully factorial experiment. Each study should display tested and untested cells and explain why selected combinations were prioritised. The benchmark dimensions and recommended reporting levels are summarised in Table 3.

Figure 3 visualises how target configuration, operating conditions and sensor state should converge on an explicit condition-stratified coverage map.

6. Core Outcome Set and Equations

6.1. Time-to-First-Detection

Let $t_0$ denote the first time at which the target is eligible for detection according to the study protocol, and $t_{\mathrm{d}\mathrm{e}\mathrm{t}}$ the time of the first valid detection that satisfies the stated confidence and persistence rule:

$$T_{\mathrm{T}\mathrm{F}\mathrm{D}}=t_{\mathrm{d}\mathrm{e}\mathrm{t}}-t_0.$$

Eligibility must be defined prospectively. It may be based on the field of view, minimum visible area, range or a known simulation onset. The definition should not be chosen after reviewing the model output.

6.2. End-to-End Latency

The total detection-to-brake-force delay should be decomposed as:

$${\tau }_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}=T_{\mathrm{T}\mathrm{F}\mathrm{D}}+L_{\mathrm{i}\mathrm{n}\mathrm{f}}+L_{\mathrm{d}\mathrm{e}\mathrm{c}}+L_{\mathrm{a}\mathrm{c}\mathrm{t}},$$

where $L_{\mathrm{i}\mathrm{n}\mathrm{f}}$ is inference and pre/post-processing latency, $L_{\mathrm{d}\mathrm{e}\mathrm{c}}$ is temporal confirmation and decision latency, and $L_{\mathrm{a}\mathrm{c}\mathrm{t}}$ is communication, actuator and brake-force build-up delay. Studies may use a more detailed decomposition, but all included and excluded components should be explicit.

6.3. Time to Collision

For a stationary target and constant longitudinal relative speed $v_{\mathrm{r}\mathrm{e}\mathrm{l}}>0$:

$$TTC={\displaystyle \frac{d}{v_{\mathrm{r}\mathrm{e}\mathrm{l}}}}.$$

This simple definition is not appropriate when relative speed, curvature or target motion changes materially; the model used in those cases should be stated.

6.4. Required Stopping Distance

Under constant effective deceleration magnitude $a_{\mathrm{e}\mathrm{f}\mathrm{f}}>0$ after the total delay:

$$d_{\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{p}}=v_0{\tau }_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}+{\displaystyle \frac{v_0^2}{2a_{\mathrm{e}\mathrm{f}\mathrm{f}}}}.$$

The first term is the distance travelled before effective braking, and the second is the idealised braking distance. If brake build-up is modelled explicitly rather than absorbed into ${\tau }_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}$ or $a_{\mathrm{e}\mathrm{f}\mathrm{f}}$, the resulting piecewise formulation should be provided.

6.5. Stopping Margin

$$M_{\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{p}}=d_{\mathrm{d}\mathrm{e}\mathrm{t}}-d_{\mathrm{s}\mathrm{t}\mathrm{o}\mathrm{p}}.$$

A positive margin indicates physical stopping feasibility under the stated assumptions. It does not prove collision avoidance because steering, road friction, target motion, controller constraints and surrounding traffic may alter the outcome.

6.6. Residual Impact Speed

When the available post-delay distance is insufficient for a full stop, an idealised residual impact speed is:

$$v_{\mathrm{i}\mathrm{m}\mathrm{p}}=\sqrt{\mathrm{m}\mathrm{a}\mathrm{x}\left\{0,v_0^2-2a_{\mathrm{e}\mathrm{f}\mathrm{f}}\left(d_{\mathrm{d}\mathrm{e}\mathrm{t}}-v_0{\tau }_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}\right)\right\}}.$$

The equation is valid only when the available braking distance $d_{\mathrm{d}\mathrm{e}\mathrm{t}}-v_0{\tau }_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}$ is interpreted consistently, and constant effective deceleration is a reasonable approximation.

6.7. Expected Calibration Error

For $K$ confidence bins $B_k$, empirical accuracy $\mathrm{a}\mathrm{c}\mathrm{c}\left(B_k\right)$ and mean confidence $\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{f}\left(B_k\right)$:

$$ECE={\displaystyle \sum _{k=1}^K}{\displaystyle \frac{\left|B_k\right|}{n}}\left|\mathrm{a}\mathrm{c}\mathrm{c}\left(B_k\right)-\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{f}\left(B_k\right)\right|.$$

ECE should be accompanied by binning rules, sample counts and preferably a reliability diagram. Adaptive or class-conditional alternatives may be appropriate when sample distributions are highly imbalanced.

6.8. Safety-Critical False-Negative Rate

Let $I_i=1$ when event $i$ is not validly detected before a predefined intervention boundary and $I_i=0$ otherwise. For $N$ eligible events:

$$FN{R}_{\mathrm{S}\mathrm{C}}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{I}_i.$$

The intervention boundary may be defined by TTC, stopping distance or a validated control criterion. It must be specified before outcome analysis. Table 4 consolidates the proposed core outcome set and the minimum reporting expectations for each outcome.

7. Reporting Analysis and Uncertainty

7.1. Unit of Analysis

The study should distinguish frame-level, sequence-level and event-level analyses. Frame-level confidence intervals are inappropriate when thousands of correlated frames arise from a small number of independent events. The preferred unit for safety claims is the independent event or scenario replicate.

When uncertainty is estimated by bootstrap, resampling should occur at the highest relevant clustering level, such as event, actor or route. The number of independent clusters and bootstrap replicates should be reported.

7.2. Threshold Selection

Confidence and persistence thresholds should be selected using training or validation data, not the final test set. Threshold sensitivity should be shown when the main conclusion changes materially across plausible settings. A single threshold optimised for the overall F1 score may not be suitable for a safety-critical rare class.

7.3. Missing and Failed Sensor Data

Studies should distinguish data that are absent because a modality was intentionally removed, technically unavailable, corrupted, outside the range or rejected by quality control. Missing data should not be silently replaced by nominal values. The fusion strategy used under missing-modality conditions should be reported and evaluated.

7.4. Statistical Reporting

Principal outcomes should include uncertainty intervals rather than only point estimates. Exact binomial, Wilson or bootstrap intervals may be used as appropriate. Multiple hypothesis testing should be controlled when many subgroup comparisons are presented. Effect sizes and absolute error counts are more informative than isolated p-values. Versioned datasets, models and evaluation artefacts can also reduce the hidden technical debt that otherwise accumulates in machine-learning systems [40].

8. Relationship to Standards, Regulations and Assessment Protocols

NUP-REPORT is intended to organise research evidence that may later inform assurance or protocol development. It does not establish compliance with any standard. Table 5 maps NUP-REPORT to selected standards, regulations and assessment frameworks while clarifying the boundary of the present proposal.

8.1. Functional Failure Versus Performance Insufficiency

A missing camera frame, corrupted sensor stream or failed communication link may be treated as a malfunctioning behaviour relevant to functional safety. A correctly functioning perception stack that fails to recognise a prone person because the target lies outside its training distribution may instead represent a functional insufficiency relevant to SOTIF. NUP-REPORT requires both categories to be reported where applicable, rather than pooling them as generic “errors”.

8.2. Regulatory and Consumer-Test Boundaries

Public vehicle-test programmes necessarily use defined, reproducible targets and scenarios. Research on non-upright postures should therefore avoid implying that an exploratory simulation is equivalent to a regulatory test. The appropriate pathway is staged: transparent reporting, representative data, target development, inter-laboratory reproducibility, threshold research and stakeholder review.

9. Recommended Implementation Pathway

Version 1.0 should be treated as a starting point. A credible path to broader adoption comprises six stages:

1. Framework release: publish the checklist, metric definitions and scenario vocabulary in reusable formats.

2. Dataset audit: quantify the prevalence and quality of non-upright annotations in existing datasets.

3. Open benchmark: release reference scenarios, hard negatives, sequence-level partitions and benchmark cards.

4. Inter-laboratory study: evaluate common scenarios across different sensors, models, laboratories and geographic settings.

5. Threshold research: link perception timing to measured vehicle response and validated injury outcomes.

6. Standards engagement: submit the accumulated evidence to consumer-test, regulatory and ISO stakeholders through established change processes.

Figure 4 presents the proposed staged pathway from framework release to standards engagement.

9.1. Versioning

Future releases should use semantic versioning:

• x: clarification or addition of reporting items without changing the conceptual framework;
• 2.0: major restructuring after external consultation or validation; and
• domain extensions: optional modules for heavy vehicles, low-speed manoeuvring, vulnerable children, wheelchair users or other specialised contexts.

Each version should include a change log and an archived checklist. Publications should cite the version applied.

10. Limitations

NUP-REPORT 1.0 has several important limitations.

First, it is a single-author framework based on structured narrative synthesis rather than a formal systematic review, Delphi process or standards committee. The proposal may therefore omit variables considered essential by other disciplines or regions.

Second, the minimum scenario matrix is intentionally broad and does not define exact parameter values. Appropriate speeds, ranges, illumination levels and degradation severities will depend on the ODD, vehicle class, sensor configuration and intended use.

Third, the framework does not provide validated performance thresholds. Defining acceptable recall, latency, calibration or stopping margin requires representative data, risk analysis and empirical vehicle testing.

Fourth, the stopping equations are simplified. Real braking and steering response may depend on tyre-road friction, brake build-up, anti-lock control, road gradient, vehicle load, curvature, target motion and surrounding traffic. Studies should use higher-fidelity models when these factors are material.

Fifth, reporting quality cannot substitute for dataset representativeness. Rare non-upright events remain difficult to capture ethically and at scale. Synthetic data and surrogates will remain necessary, but their domain gap must be quantified.

Finally, the framework cites prior work by the author and collaborators because those publications directly motivate the detection, uncertainty and safety-integration problem [14,15,16,17]. The present article does not reuse their datasets, case material or unpublished contributions; it proposes an independent reporting structure based on published literature and public standards.

11. Conclusions

Non-upright pedestrians present a distinct perception and pre-crash safety-evaluation problem. Studies cannot be compared reliably when posture labels, target surrogates, data partitions, environmental conditions, sensor states, latency boundaries and vehicle-response assumptions remain implicit.

NUP-REPORT 1.0 provides a six-domain framework covering target and posture; scenario and environment; sensors and data provenance; model and fusion architecture; performance and uncertainty; and vehicle-level safety interpretation. It adds a minimum scenario matrix, core outcome set, explicit timing and stopping-margin equations, standards mapping and a 30-item checklist.

The central principle is that object-detection accuracy is not equivalent to safety performance. A detector must be evaluated in terms of when and where it produces a valid output, how its confidence behaves under shift, how it degrades when sensors fail, and what physical intervention opportunity remains. By making these elements visible, NUP-REPORT can support more reproducible research, more meaningful comparison and a defensible path from exploratory studies to shared benchmarks and, eventually, evidence-based protocol development.