IRIS: Infrastructure-aware Multimodal Reasoning for Road Safety Auditing with Regulatory Compliance

We introduce IRIS, a multimodal framework for infrastructure-oriented road-safety auditing that combines synchronized egocentric video, inertial measurements, GPS, and a machine-readable Indian Roads Congress (IRC) knowledge graph. IRIS treats localized evasive events as complementary evidence, reasoning about their associated infrastructure conditions, locations, and compliance with applicable IRC requirements.

Data Collection

IRIS is collected using a low-cost mobile sensing platform deployed on a two-wheeler during real-world road traversals.

Sensor	Data	Rate
Forward-facing camera (GoPro)	Egocentric video	30 FPS
Smartphone IMU (Phyphox)	Tri-axial acceleration and angular velocity	200 Hz
Smartphone GPS	Latitude, longitude, altitude and speed	1 Hz

This release contains 4 recording sessions (S1 to S4), about 25 GB in total.

Structure

S1/ ... S4/                              # one folder per recording session
├── GX01xxxx.MP4                         # GoPro video, part 1
├── GX02xxxx.MP4                         # GoPro video, part 2 (continuation)
└── phyphox/
    ├── Accelerometer.csv                # accelerometer incl. gravity
    ├── Gyroscope.csv                    # angular velocity
    ├── Linear Acceleration.csv          # gravity removed 
    └── Location_R<k>.csv                # GPS

Session	Video files	Total size
S1	GX019990, GX029990	~5.4 GB
S2	GX019991, GX029991	~5.0 GB
S3	GX019993, GX029993	~7.1 GB
S4	GX019994, GX029994	~6.2 GB

Telemetry CSV schema

All CSVs use a Time (s) column (seconds since the Phyphox experiment started) and scientific notation.

File	Columns
Accelerometer.csv	Time (s), Acceleration x/y/z (m/s^2) (includes gravity)
Gyroscope.csv	Time (s), Gyroscope x/y/z (rad/s)
Linear Acceleration.csv	Time (s), Linear Acceleration x/y/z (m/s^2) (gravity removed)
Location_R<k>.csv	Time (s), Latitude, Longitude, Height (m), Velocity (m/s), Direction, Horizontal Accuracy (m), Vertical Accuracy (m) (may contain NaN)

IMU channels are sampled at 200 Hz (about 5 ms between samples) and GPS at 1 Hz.

Sample Usage

Download the whole dataset:

from huggingface_hub import snapshot_download
import pandas as pd

root = snapshot_download("IRIS-Roadsafety/IRIS", repo_type="dataset")
acc = pd.read_csv(f"{root}/S1/phyphox/Accelerometer.csv")

Download only the telemetry (skip the large videos):

snapshot_download("IRIS-Roadsafety/IRIS", repo_type="dataset",
                  allow_patterns=["*/phyphox/*"])

Notes

• Location files contain an encoding artefact in the column headers (°); read with encoding="utf-8" or rename the columns.
• Sensor timestamps are relative to the start of the Phyphox recording, not to the start of the video; align the two using a common event.

Citation

@inproceedings{iris2026,
  title={IRIS: Infrastructure-aware Multimodal Reasoning for Road Safety Auditing with Regulatory Compliance},
  author={Vedula, Abhishek and Keerthana, A R and Nidamanuri, Jaswanth and Subramanian, Anbumani and Jawahar, C V},
  booktitle={Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP)},
  year={2026}
}

License

Released under CC BY-NC 4.0: non-commercial research use only, with attribution.