A physics simulation and statistical study of lateral intrusion detection gaps
in production ADAS sensor architectures at highway speeds
Phase 2 of the ADAS Validation Gap Research Series
Overview · Research Journey · Key Findings · Methodology · Results · Robustness Validation · How to Run · Citation · About
| Motorcycle — Chase Camera | Motorcycle — Overhead View | Passenger Car — Chase Camera |
|---|---|---|
 |
 |
 |
From left: CARLA simulation confirms MC is visible but undetected → sensor architecture explains why → Monte Carlo quantifies 99.6% failure rate → counterfactual maps the solution space
| Baseline collision rate 100,000 scenarios |
All sensors detect 0.01 s after departure |
AEB activates TTC = 3.43 s threshold |
Reaction window left Insufficient at 80 km/h |
| Corner radar improvement 99.6% → 77.4% |
Physics-limited floor Unavoidable even with perfect detection |
Real-world severe injury 1.53M German records 2.37× national baseline |
Euro NCAP covers ELK (departing vehicle only). AEB for receiving vehicle: no protocol exists. |
Euro NCAP update — October 2025: The Lane Departure Collision Protocol v1.1 introduces Emergency Lane Keeping (ELK) assessment for Car-to-Motorcyclist scenarios — testing steering correction for the departing vehicle. It does not test AEB response for the receiving vehicle when a lateral intruder enters its path. The gap this study identifies remains unaddressed in active safety validation frameworks.
Modern AEB systems are designed and certified for longitudinal threats — vehicles braking ahead of you. Euro NCAP tests rear stationary, rear moving, pedestrian, and cyclist scenarios. Every single one is from the front. In a straight line.
None test whether the receiving vehicle's AEB can stop a lateral intruder.
This study asks one question:
If a motorcycle loses control and crosses into your lane laterally at highway speed — can current production AEB detect and respond in time?
No manufacturer demonstrates this scenario. No Euro NCAP AEB protocol tests it. This study builds the physics simulation, runs 100,000 randomised scenarios, and finds the answer is no — at two independent architectural levels.
This is not a standalone simulation. It is the direct continuation of a question that started with 1.53 million crash records.
╔══════════════════════════════════════════════════════════════════════╗
║ PHASE 1 — adas-validation-gap-analysis ║
║ github.com/TejasManjunath/adas-validation-gap-analysis ║
╠══════════════════════════════════════════════════════════════════════╣
║ ║
║ Data: 1,539,249 German crash records, 2019–2024 ║
║ Source: Unfallatlas — Statistische Ämter des Bundes ║
║ und der Länder. Licence: Data Licence Germany v2.0 ║
║ ║
║ Method: Logistic regression severity model + SRPI ranking ║
║ Cross-referenced against Euro NCAP AEB protocol documents ║
║ ║
║ Finding: Motorcycle leaving-carriageway = 44.1% severe injury rate ║
║ 2.37× the national baseline of 18.56% ║
║ Zero coverage in Euro NCAP AEB test protocols ║
║ ║
╚══════════════════════════════════════════════════════════════════════╝
│
│ The gap is proven in crash data.
│ But can current hardware handle it?
│
▼
╔══════════════════════════════════════════════════════════════════════╗
║ PHASE 2 — aeb-blind-side-study (THIS REPOSITORY) ║
╠══════════════════════════════════════════════════════════════════════╣
║ ║
║ Tool: CARLA 0.9.16 physics simulation ║
║ Scale: 100,000 Monte Carlo scenarios · multi-seed validated ║
║ ║
║ Finding: Detection occurs at t = 5.01 s — 0.01 s after departure. ║
║ The system sees it. Then waits 3 more seconds. ║
║ AEB activates at t = 8.05 s — 1.1 s remains. ║
║ Not missing hardware. Architectural mismatch. ║
║ ║
╚══════════════════════════════════════════════════════════════════════╝
Phase 1 proved the gap is real using crash statistics.
Phase 2 proves the gap is structural using physics simulation.
Together: a complete evidence chain from injury data to root cause.
Figure 1 — Phase 1 finding from 1.53M German crash records: motorcycle leaving-carriageway scenarios show 44.1% severe injury rate — 2.37× the national baseline. Zero Euro NCAP AEB coverage. This is the gap Phase 2 investigates.
Figure 2 — At t = 5.0 s: motorcycle clearly visible at 3.5 m lateral separation. Forward radar (35°): NO DETECTION — structural FOV gap. Side radar (8 m): NON-ACTIONABLE — parking-assist spec (<10 km/h). Rear radar: NO DETECTION — geometry gap. AEB has no trigger pathway.
Figure 3 — Left: current production architecture — 35° forward FOV blind zone, parking-spec side radar. Right: corner radar enhancement (120° FOV, 50 m) restores lateral coverage. Collision rate: 99.6% → 77.4% (−22.1 pp).
Two independent architectural failures compound each other. Fixing one without the other produces minimal improvement — as the counterfactual results confirm.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
LAYER 1 — SENSOR ARCHITECTURE (why the system cannot act in time)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Forward radar 35° FOV — MC is geometrically always outside the cone
Side radar 8 m, <10 km/h spec — non-actionable at highway speed
Rear radar MC approaches laterally — first valid hit at t = 13.3 s
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
LAYER 2 — DECISION ARCHITECTURE (why detection alone is not enough)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
t = 5.01 s All sensors detect — 0.01 s after departure
The system is not blind. Detection is not the failure.
t = 5–8 s TTC threshold not crossed — AEB waits
TTC logic built for longitudinal threats
Lateral closing at 0.70 m/s drops TTC 2.5× slower
than a typical longitudinal approach at 80 km/h
t = 8.05 s AEB activates — TTC = 3.43 s threshold finally crossed
t = 9.15 s Collision — only 1.1 s was available
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
CONCLUSION: Not missing hardware. A specification mismatch and a
decision architecture calibrated for the wrong geometry.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
aeb-blind-side-study/
│
├── 📄 README.md
├── 📄 LICENSE
├── 📄 requirements.txt
├── 📄 .gitignore
│
├── 📁 Simulation/
│ ├── s3_ego_mc.py ← CARLA S1: EGO + Motorcycle (baseline)
│ ├── s3_ego_mc_v2.py ← CARLA S1 v2: + CF sweep + TTC sweep modes
│ ├── s3_ego_car.py ← CARLA S2: EGO + Passenger Car
│ └── s4_analysis.py ← Detection window figures vs Euro NCAP
│
├── 📁 Analysis/
│ └── monte_carlo_final.py ← MC: baseline + counterfactual + actuation sweep
│
├── 📁 figures/
│ ├── mc_incident.gif ← MC lateral departure — chase camera
│ ├── mc_overhead.gif ← MC lateral departure — overhead view
│ ├── car_incident.gif ← Car lateral departure — chase camera
│ ├── hero_figure.png ← Four-panel research summary
│ ├── 08_validation_story_visual.png← Phase 1 crash severity chart
│ ├── figA_simulation_storyboard.png
│ ├── figB_sensor_gap_proof.png
│ ├── figC_overhead_sequence.png
│ ├── figD_car_storyboard.png
│ ├── figE_car_overhead.png
│ ├── figF_mc_vs_car_comparison.png
│ ├── figG_aeb_timeline.png
│ ├── sensor_architecture_RIGHT.png
│ ├── sensor_coverage_diagram.png
│ ├── annotated_scenario_forensic.png
│ ├── time_interaction_model_FINAL.png
│ ├── event_timeline.png
│ ├── detection_window_chart.png
│ ├── detection_window_table.png
│ ├── failure_attribution.png
│ ├── cf_comparison.png
│ ├── mc_distribution.png
│ ├── mc_heatmap.png
│ ├── actuation_heatmap.png
│ ├── impact_severity.png
│ ├── mc_headline.png
│ ├── fig_s3_correction_derivation.png ← S3 = (3.50−0.60)/3.50 derivation
│ ├── fig_cf_sweep_validation.png ← CF validation: mean=0.833, σ=0.005
│ ├── fig_ttc_sweep_results.png ← 8 TTC thresholds, all collision YES
│ ├── fig_multiseed_validation.png ← Seeds 42/123/999, std=0.020%
│ └── fig_sensitivity_analysis.png ← 3-panel sensitivity bars
│
├── 📁 frames/
│ ├── mc/ ← CARLA frames — MC scenario
│ └── car/ ← CARLA frames — Car scenario
│
└── 📁 data/
├── sensor_logs/
│ ├── mc_sensor_log.txt
│ ├── car_sensor_log.txt
│ ├── mc_sensor_raw.txt
│ └── car_sensor_raw.txt
└── outputs/
├── mc_baseline_metrics.csv
├── mc_tick_data.csv
├── car_baseline_metrics.csv
├── car_tick_data.csv
├── cf_summary.csv
├── cf_sweep_results.csv
├── actuation_sweep.csv
└── ttc_sweep_results.csv
The scenario replicates the crash topology identified in Phase 1: a vehicle departing laterally into an adjacent lane at highway speed. Town04's straight highway section was chosen deliberately to isolate sensor and decision architecture performance without curve geometry confounds.
| Parameter | Value | Rationale |
|---|---|---|
| Simulator | CARLA 0.9.16 | Physics-validated, open-source ADAS platform |
| Map | Town04 — straight highway | Eliminates curve geometry as confounding variable |
| Ego vehicle | BMW Grand Tourer | Representative mid-size crossover |
| Threat S1 | Kawasaki Ninja | Phase 1 SRPI top-ranked scenario |
| Threat S2 | Toyota Prius | Target-type robustness validation |
| Speed — both | 80 km/h constant | Highway speed at departure conditions |
| Lane gap | 3.50 m | Measured from Town04 waypoint geometry |
| Lateral drift | 0.70 m/s | Representative loss-of-control departure rate |
| Correction factor | 0.83 | Kinematic threshold ratio — see derivation below |
| Collision window | 4.15 s (CARLA) | vs 5.00 s mathematical prediction |
| OS / Python | Windows 11 / Python 3.10 | Ryzen 7, RTX 5060 Mobile, 32 GB RAM |
Figure 4 — Chase camera: stable travel t = 0–5 s → lateral departure t = 5–9 s (zero actionable detection) → unavoidable collision t = 9.15 s. AEB was never activated.
Overhead view — Motorcycle crossing lane markings laterally into ego vehicle path. The geometric blind zone of the 35° forward radar is clear from this perspective.
The analytical crossing-time model predicts t_cross = 3.50 / 0.70 = 5.00 s. CARLA physics measures 4.15 s. The discrepancy arises because CARLA's collision detection triggers at a finite lateral separation, not at zero.
The correction factor is not an empirical fit — it is a kinematic threshold ratio derived from first principles:
S3 = (LANE_GAP − IMPACT_THRESH) / LANE_GAP
= (3.50 − 0.60) / 3.50
= 0.83
Where:
LANE_GAP = 3.50 m (Town04 measured)
IMPACT_THRESH = 0.60 m (motorcycle body half-width at physics handoff)
Applied globally: time_to_cross = (gap / drift) × 0.83
Validated across five drift speeds (0.3–1.2 m/s): mean = 0.833, σ = 0.005, CV = 0.56%. Drift-independent — confirming it is a geometric property, not a speed-dependent calibration.
Figure 5 — Kinematic threshold ratio derivation: (3.50 m − 0.60 m) / 3.50 m = 0.83. IMPACT_THRESH = 0.60 m represents the lateral separation at which CARLA physics collision detection triggers.
| Sensor | H-FOV | Range | Production Spec | Finding |
|---|---|---|---|---|
| Forward radar | 35° | 150 m | Standard AEB LRR | FOV gap — lateral targets always outside cone |
| Side radar | 120° | 8 m | Parking-assist SRR | Non-actionable — <10 km/h design threshold |
| Rear radar | 150° | 80 m | BSD / RCTA | Geometry gap — first valid hit at t = 13.3 s |
| Corner radar | 120° | 50 m | Counterfactual only | Restores coverage — −22.1 pp improvement |
Modelling constraint: CARLA uses geometric ray-cast detection — no radar cross-section (RCS), multipath, or clutter simulation. This represents an upper bound on detection capability. Real motorcycle RCS (~1–3 m²) is significantly lower than cars (~10–20 m²), meaning real-world failure rates would be equal to or worse than the 99.6% baseline reported here.
Figure 6 — Radar coverage during stable phase (left) and departure phase (right). All sensor modalities fail during the 5.0 s departure window. Zero seconds detectable vs 1.5 s minimum required for AEB intervention.
The Monte Carlo framework evaluates a closed-form analytical collision model calibrated from CARLA scenario observations — not 100,000 independent CARLA physics runs. This enables 3–5 minute execution on standard hardware while preserving physics fidelity.
Model equations:
t_cross = (gap / drift) × 0.83
t_respond = reaction_time + (speed / deceleration)
collision = t_cross ≤ t_respond OR detection = False
| Parameter | Distribution | Range | Note |
|---|---|---|---|
| Speed | Stratified uniform | 60–100 km/h | 60% weight in 70–90 km/h highway band |
| Lateral drift | Uniform | 0.3–1.2 m/s | Gradual drift to rapid loss-of-control |
| Lane gap | Uniform | 2.5–5.0 m | European highway lane width variation |
| Longitudinal offset | Uniform | ±10 m | Pure lateral departure geometry (excludes cut-ins) |
| Seeds | — | 42, 123, 999 | Std = 0.020% |
On the ±10 m offset: Wider offset ranges (e.g. −30 to +50 m) introduce cut-in scenarios from trailing positions, artificially inflating AEB viability rates. The ±10 m range is constrained deliberately to model pure lateral departure geometry only.
Figure 7 — Euro NCAP scenarios provide 1.9–3.8 s of actionable detection window. Lateral intrusion: zero actionable window. The Phase 1 protocol gap (no Euro NCAP coverage) is confirmed as a Phase 2 physics gap (zero AEB viability).
Production sensor architecture produced collision outcomes in 99.57% of 100,000 Monte Carlo scenarios (reported as 99.6%). Multi-seed validation confirms stability: seeds 42, 123, 999 yielded 99.57%, 99.59%, 99.54% (std = 0.020%).
Figure 8 — Three independent random seeds produce collision rates within 0.05 pp of each other (std = 0.020%). The baseline finding is not seed-dependent.
Figure 9 — Drift × gap parameter space: realistic highway conditions (0.7–1.2 m/s, 3–5 m) fall entirely in the collision zone. The 0.4% AEB-viable region exists only at extreme slow-drift and large-gap boundaries — outside realistic highway departure conditions.
Figure 10 — 53.2% of failures are detection-limited (addressable through sensor architecture improvement). 46.6% are physics-limited — unavoidable at 80 km/h even with perfect detection and best-case braking.
| Configuration | Collision Rate | vs Baseline | Key Change |
|---|---|---|---|
| Baseline — production | 99.57% | — | 35° fwd FOV gap + parking-spec side radar |
| Improved side radar | 97.38% | −2.19 pp | Highway threshold, 8 m range retained |
| Corner radar (120°/50 m) | 77.42% | −22.15 pp | FOV + range restored at front corners |
| Perfect detection | 46.62% | −52.95 pp | All sensor constraints removed |
Figure 11 — Corner radar delivers 10.1× larger improvement than side radar threshold upgrade alone (−22.15 pp vs −2.19 pp). Range, not speed threshold, is the binding constraint for the side radar. Corner sensors address both FOV coverage and detection range simultaneously.
Under perfect detection, a 4×3 grid of braking parameters was evaluated. Best-case (0.5 s reaction, 10 m/s²) achieves 16.48% — approximately 1 in 6 scenarios still ends in collision. Longitudinal braking reduces ego speed but cannot stop lateral closing motion.
Figure 12 — Actuation parameter sweep under perfect detection. Even the best combination (0.5 s, 10 m/s²) leaves 16.48% unavoidable — confirming the physics-limited floor.
Figure 13 — Mean impact speed: 22.77 m/s (82 km/h) — near full speed in almost every scenario. No AEB intervention means no speed reduction at impact. This directly explains the 44.1% severe injury rate found in Phase 1.
Three independent checks confirm the 99.6% baseline is not an artefact of parameter selection.
Five independent CARLA runs at different drift speeds validate that the 0.83 kinematic threshold ratio is drift-independent.
| Drift (m/s) | Predicted (s) | CARLA Actual (s) | Ratio |
|---|---|---|---|
| 0.3 | 11.67 | 9.70 | 0.8314 |
| 0.5 | 7.00 | 5.80 | 0.8286 |
| 0.7 | 5.00 | 4.15 | 0.8300 |
| 0.9 | 3.89 | 3.25 | 0.8357 |
| 1.2 | 2.92 | 2.45 | 0.8400 |
| Mean | 0.8331 | ||
| Std | 0.0047 |
CV = 0.56% — consistent across the full drift range. The factor is geometric, not speed-dependent.
Figure 14 — Correction factor across five drift speeds: mean = 0.8331, σ = 0.0047, CV = 0.56%. The ratio is drift-independent, confirming it is a geometric property of the simulation environment.
Eight TTC activation thresholds spanning the full realistic production specification range were tested. All eight produced collision outcomes.
| TTC Threshold (s) | AEB Triggered | AEB Active (s) | Window (s) | Collision |
|---|---|---|---|---|
| 1.50 | No | — | — | YES |
| 2.00 | Yes | 11.35 | −2.20 | YES |
| 2.50 | Yes | 12.60 | −3.45 | YES |
| 3.00 | Yes | 10.45 | −1.30 | YES |
| 3.43 | Yes | 9.00 | +0.15 | YES |
| 4.00 | Yes | 7.60 | +1.55 | YES |
| 4.50 | Yes | 6.80 | +2.35 | YES |
| 5.00 | Yes | 6.05 | +3.10 | YES |
Two failure regimes: For thresholds below 3.43 s, the minimum TTC reached during the 4.15 s departure window is 3.34 s — AEB never fires or fires post-collision. For thresholds above 3.43 s, AEB fires earlier but longitudinal braking cannot stop lateral closing motion regardless of activation time.
The 99.6% result is not a threshold calibration artefact. The failure is structural.
Figure 15 — All eight TTC thresholds (1.5–5.0 s) produce collision outcomes. The finding is threshold-independent across the full realistic production specification range.
Figure 16 — Sensitivity analysis: collision rate across speed bands (60–100 km/h), drift bands (0.3–1.2 m/s), and gap bands (2.5–5.0 m). High collision rates persist across all parameter ranges — the 99.6% baseline is not concentrated in any single parameter region.
── MULTI-SEED BASELINE VALIDATION ──
Seed 42: 99.57%
Seed 123: 99.59%
Seed 999: 99.54%
Mean: 99.57% Std: 0.020%
Target 99.6% ±0.5%: PASS ✓
The passenger car scenario (Toyota Prius, S2) was run under identical parameters (80 km/h, 3.50 m gap, 0.70 m/s drift) to test whether the failure depends on motorcycle-specific radar cross-section.
It does not.
| Parameter | Motorcycle (S1) | Passenger Car (S2) |
|---|---|---|
| Detection latency | 0.011 s | 0.002 s |
| First detection | t = 5.011 s | t = 5.002 s |
| AEB activation | t = 8.05 s | t = 8.05 s |
| Collision | t = 9.15 s | t = 9.15 s |
| Response window | 1.10 s | 1.10 s |
| Outcome | COLLISION | COLLISION |
The car was detected 0.009 s faster (larger geometric cross-section in ray-cast model). Every critical timing was identical. The failure mechanism is architectural, not target-specific. Within CARLA's geometric sensor model, both target types yield the same system-level outcome.
Note: This validates internal simulation consistency. Real-world motorcycle detection would be significantly harder due to lower RCS (~1–3 m²) compared to cars (~10–20 m²). The car scenario does not imply real-world RCS equivalence.
Figure 16 — Motorcycle and passenger car produce identical system-level timing: detection t = 5.01 s · AEB t = 8.05 s · collision t = 9.15 s. The failure is in decision architecture, not target detectability.
📎 Additional Simulation Figures
Figure 17 — Event timeline overlay: both S1 and S2 produce identical phase transitions. Detection window (5.01–8.05 s) and response window (1.10 s) are unchanged across target types.
Figure 18 — Passenger car scenario chase camera: stable phase → departure → collision at t = 9.15 s. Detection depths at departure: fwd 14.6 m · side 2.6 m · rear 10.0 m.
Figure 19 — Forensic top-down diagram: motorcycle departure trajectory outside the 35° forward radar cone. Side radar non-actionable above 10 km/h. No AEB trigger pathway exists in production architecture.
| # | Assumption | Scope Impact | Notes |
|---|---|---|---|
| 1 | Geometric detection — no RCS modelling | Upper bound | Real motorcycle detection would be harder; real failure rates ≥ 99.6% |
| 2 | Single CARLA speed (80 km/h) | Constrained | Monte Carlo extends analytically to 60–100 km/h |
| 3 | Uniform parameter distributions | Standard | Common approach in safety scenario analysis |
| 4 | S3 = 0.83 applied globally | Validated | Confirmed across 5 drift speeds (CV = 0.56%); gap variation not tested in CARLA |
| 5 | Straight road only | Constrained | Isolates sensor/decision architecture — curve geometry is future work |
| 6 | Constant lateral drift | Slightly conservative | Real loss-of-control events may exhibit accelerating drift |
| 7 | No evasive steering | Constrained | Isolates longitudinal braking capability only |
| 8 | Linear braking model | Constrained | Not full CARLA vehicle dynamics; isolates AEB timing analysis |
Interpretation boundary: The 99.6% finding represents parameter space coverage under modelled conditions — the architecture cannot handle this scenario topology when it occurs under these assumptions. It is not a real-world crash probability estimate. Production validation under diverse conditions and ISO 21448 (SOTIF) compliance verification are required before these findings inform any design or regulatory decision.
pip install -r requirements.txtCARLA 0.9.16 server must be running on
localhost:2000
See CARLA quickstart
# Scenario S1 — Motorcycle lateral departure (baseline)
python Simulation/s3_ego_mc.py
# Scenario S2 — Passenger car lateral departure
python Simulation/s3_ego_car.pyOutput files per run
mc_sensor_log.txt — departure onset, AEB timing, collision event summary
mc_sensor_raw.txt — raw radar returns per sensor per tick
mc_baseline_metrics.csv — kinematic calibration data for Monte Carlo
mc_tick_data.csv — per-tick position, TTC, lateral separation
cam_chase_t*.png — chase camera frames (when SAVE_FRAMES=True)
cam_overhead_t*.png — overhead frames at 18 m altitude
cam_side_t*.png — right-side camera frames
s3_ego_mc_v2.py extends the baseline script with two sweep modes. Set the flag at the top of the file before running.
# In s3_ego_mc_v2.py — set exactly one to True:
CF_SWEEP_MODE = True # Correction factor validation across 5 drift speeds
TTC_SWEEP_MODE = False # TTC threshold sensitivity across 8 thresholdspython Simulation/s3_ego_mc_v2.pyExpected outputs:
# CF sweep: cf_sweep_results.csv — mean=0.8331, σ=0.0047
# TTC sweep: ttc_sweep_results.csv — 8/8 thresholds → collision YES
# Baseline + counterfactual + actuation sweep (~3–5 min)
python Analysis/monte_carlo_final.pyExpected output — verify your run matches:
── MULTI-SEED BASELINE VALIDATION ──
Seed 42: 99.57%
Seed 123: 99.59%
Seed 999: 99.54%
Mean: 99.57% Std: 0.020%
Target 99.6% ±0.5%: PASS ✓
Output files
cf_comparison.png — four-bar counterfactual architecture comparison
cf_summary.csv — collision rates per sensor configuration
actuation_heatmap.png — reaction time × deceleration sweep
actuation_sweep.csv — actuation sweep raw results
mc_headline.png — 99.6% headline statistic
mc_distribution.png — crossing time histogram
mc_heatmap.png — drift × gap parameter space scatter
failure_attribution.png — pie + bar failure attribution
impact_severity.png — impact speed distribution
python Simulation/s4_analysis.py@misc{manjunath2026aeb,
author = {Tejas Manjunath},
title = {AEB Blind Side Study: Lateral Intrusion ADAS Gap —
Physics Simulation and Monte Carlo Analysis},
year = {2026},
publisher = {GitHub / Zenodo},
url = {https://github.com/TejasManjunath/aeb-blind-side-study},
doi = {10.5281/zenodo.XXXXXXX},
note = {Phase 2 of the ADAS Validation Gap Research Series}
}Phase 1 (prerequisite reading):
@misc{manjunath2026adas,
author = {Tejas Manjunath},
title = {ADAS Scenario Coverage Gap Analysis: Using 1.53M German
Crash Records to Identify Euro NCAP AEB Validation Blind Spots},
year = {2026},
publisher = {GitHub},
url = {https://github.com/TejasManjunath/adas-validation-gap-analysis}
}| # | Source | Used For |
|---|---|---|
| [1] | Dosovitskiy, A., et al. (2017). CARLA: An open urban driving simulator. CoRL 2017. https://proceedings.mlr.press/v78/dosovitskiy17a.html | CARLA simulator |
| [2] | Statistische Ämter des Bundes und der Länder. (2024). Unfallatlas [Dataset]. https://unfallatlas.statistikportal.de/ Licence: Data Licence Germany — Attribution v2.0 | Phase 1 crash data |
| [3] | Euro NCAP. (2025). Crash Avoidance Lane Departure Collisions Protocol, v1.1. Implementation Jan 2026. | ELK lateral protocol |
| [4] | Euro NCAP. (2022). Test Protocol — AEB/LSS VRU Systems, v4.3. https://cdn.euroncap.com/media/75436/euro-ncap-aeb-lss-vru-test-protocol-v43.pdf | AEB longitudinal protocols |
| [5] | Hayward, J. C. (1972). Near miss determination through use of a scale of danger. Highway Research Record, 384, 24–34. | TTC origin |
| [6] | Behera, A., et al. (2025). An improved two-dimensional time-to-collision for articulated vehicles. arXiv:2507.04184. | TTC lateral limitation |
| [7] | National Highway Traffic Safety Administration. (2024). FMVSS No. 127 — AEB Systems for Light Vehicles. Federal Register, 89 FR 39686. | NHTSA AEB mandate |
| [8] | ISO 21448:2022 — Road Vehicles: Safety of the Intended Functionality (SOTIF). | Safety validation standard |
| [9] | UNECE Regulation No. 152 — Advanced Emergency Braking System. (2021). | AEB regulation |
This study began with a dataset question, not a simulation idea.
An analysis of 1.53 million German crash records — spanning 2019 through 2024 (Unfallatlas, Statistische Ämter des Bundes und der Länder) — found motorcycle leaving-carriageway events producing severe injury rates 2.37× the national average. A Scenario Risk Priority Index (SRPI) ranked them first across all scenario types. Cross-referencing against Euro NCAP AEB documentation confirmed they were entirely absent from any published AEB validation framework.
That raised the question that became this project: not whether the scenario was tested — it was not — but whether current production AEB hardware could actually handle it if it were.
CARLA was used to build the scenario iteratively. The initial finding was unexpected: the sensors detected the motorcycle 0.01 seconds after departure. The system was not blind. The failure was that TTC-based AEB decision logic took 3 more seconds to reach its threshold — by which point only 1.1 seconds remained. Further investigation confirmed why: lateral closing speed at 0.70 m/s drops TTC 2.5× more slowly than a standard longitudinal approach, making a threshold calibrated for longitudinal threats structurally inadequate for lateral intrusion events.
TTC threshold sensitivity analysis across eight values confirmed this is not a calibration problem. All eight thresholds produce collision. Correction factor validation across five drift speeds confirmed the kinematic model is drift-independent. Monte Carlo analysis across 100,000 scenarios confirmed the result is stable across random seeds. Counterfactual analysis confirmed corner radar delivers 10× larger improvement than side radar threshold upgrades. The failure is structural at both the sensor architecture and decision logic levels.
Both phases were built independently by Tejas Manjunath between late 2025 and 2026. One laptop. No funding. No institutional support.
If you work in automotive safety and this is useful — that is why it was made public. If you use it, cite it. If you want to discuss or challenge the methodology, open an issue or reach out directly.
Code: MIT — see LICENSE
Research content, figures, and methodology: © 2026 Tejas Manjunath
Attribution required for any use of figures or findings:
Tejas Manjunath (2026). AEB Blind Side Study — Lateral Intrusion ADAS Gap.
https://github.com/TejasManjunath/aeb-blind-side-study
Phase 2 of the ADAS Validation Gap Research Series
Built with CARLA 0.9.16 · Python 3.10 · NumPy · Matplotlib · Pillow
⭐ If this work is useful, please star the repository
adas · aeb · autonomous-driving · carla-simulator · lateral-collision · motorcycle-safety · monte-carlo · sensor-architecture · safety-simulation · euro-ncap · sotif · adas-validation · data-science