Authors: Dr. Nick Barua¹ · Prof. Masahito Hitosugi² ¹ Department of Legal Medicine, Shiga University of Medical Science, Otsu, Shiga, Japan Published: Vehicles, Vol. 7, No. 4, p. 149 (2025) · Part 1 of 4 in the AFODS Research Program
Collisions with fallen pedestrians pose a lethal challenge to current Advanced Driver Assistance Systems (ADAS). Standard monocular architectures yield a True Positive Rate (TPR) as low as 21.4% at night for prone individuals — a critical classification gap.
This repository contains the software framework and quantitative evaluation for the Advanced Falling Object Detection System (AFODS). AFODS architecturally integrates LWIR Thermal, NIR Stereo, and Ultrasonic sensors with a custom AI pipeline to achieve 98.2% TPR in daytime conditions and 95.6% TPR at night.
To ensure reproducibility, the supporting data for this paper is archived on Zenodo:
- Key Figures & Diagrams: https://zenodo.org/records/14841919
- Operational Sequence Video: https://zenodo.org/records/14856943
- Archived Source Code:
AFODS integrates four layers to satisfy the redundancy and explainability requirements of ASIL C/D:
- Spatial Detection: YOLOv7-Tiny optimized for edge-case pedestrian postures (prone/falling) — mAP@0.5: 91.3%
- Predictive Kinematics: RNN + Kalman-filter state estimation — generates alerts 0.3–0.8s before ground contact
- Acoustic Verification: MFCC-based classification distinguishes fall acoustic signatures from road noise
- Explainability: SHAP values provide a forensic audit trail admissible in post-incident reconstruction
- Hypothermia Detection Mode: Thermal variance analysis between subject (36.5–37.5°C) and ambient environment
The combined detection probability
where
Using NIR stereo disparity, the system calculates vertical velocity (
where
AFODS was validated using Anthropomorphic Test Devices (ATDs) and pneumatic fall rigs across forensic-based fall archetypes:
| Condition | TPR (%) | FPR (%) | mAP@0.5 (%) | Latency (ms) |
|---|---|---|---|---|
| Daytime, Clear | 98.2 | 1.8 | 91.3 | 38 |
| Night, Dry Road | 95.6 | 3.1 | 88.7 | 42 |
| Night, Rain | 89.4 | 5.2 | 83.1 | 51 |
| Baseline (Monocular, Night) | 21.4 | N/A | N/A | N/A |
The detection of fallen pedestrians maps to ASIL C-D under ISO 26262:
- Severity (S3): Life-threatening or fatal injuries
- Exposure (E3): Significant occurrence in urban and night environments
- Controllability (C0): Near-zero ability for a driver to avoid the hazard once within the classification gap
This repository is Part 1 of a unified 4-paper road safety research program:
| # | Title | Venue | Role |
|---|---|---|---|
| 1 | Advanced Multi-Modal Sensor Fusion System (this repo) | MDPI Vehicles | Technical foundation & benchmarks |
| 2 | From Post-Mortem to Prevention: Redefining "Invisible" Pedestrians through ISO 26262 and Multi-Modal AI | SSRN | Problem framing & ISO 26262 compliance |
| 3 | Integrated Safety Architectures: Leveraging Multi-Modal AI and ISO 26262 to Protect Vulnerable Road Users | SSRN | System-level VRU architecture |
| 4 | Sudden Incapacitation or Death at the Wheel: Unravelling the Predictors of Catastrophic Multi-Vehicle Collisions | SSRN (pending) | Epidemiological evidence for ADAS mandate |
- From-Post-Mortem-to-Prevention-AFODS — ISO 26262-aligned
- AFODS-Sensor-Fusion-Code — Implementation logic
- Leveraging-Multi-Modal-AI-and-ISO-26262-to-Protect-Vulnerable-Road-Users — Safety framework
To ensure reproducibility, the supporting data for this paper is archived on Zenodo:
- Key Figures & Diagrams:
- Operational Sequence Video:
- Archived Source Code:
