deepannotate-ai/car-driving-multimodal-v1

Car Driving Multimodal Dataset v1

Overview

The Car Driving Multimodal Dataset v1 is a publication-quality egocentric dataset designed to advance research in human-object interactions (HOI), egocentric vision, and autonomous cabin activities. It captures a driver's hand interactions, hand poses, activity recognition labels, and synthetic inertial/depth annotations from a first-person perspective.

This dataset was generated using the Hand Egocentric Multimodal Annotation Pipeline, which integrates:

• Egocentric RGB-D sequence recording.
• Automated 2D and 3D hand keypoints estimation.
• Hand-object interaction mapping.
• Action and semantic natural language labeling.
• Synthetic/pseudo sensor generation (IMU signals, motion statistics, and estimated depth).

By mapping fine-grained hand pose variations alongside physical motion profiles and environment objects, this dataset serves as a benchmark for training Embodied AI and action-recognition systems in vehicle environments.

---

Dataset Highlights

• Egocentric RGB & Monocular Depth: High-resolution first-person recordings paired with frame-aligned monocular depth estimation.
• 2D/3D Hand Pose Keypoints: 21 landmark locations tracked in pixel coordinates and metric 3D space.
• Hand-Object Interactions (HOI): Direct mapping of active hand-object intersections per frame.
• Object Annotations: Detailed bounding boxes and labels for cabin objects (e.g., person, car).
• Motion Statistics & Pseudo IMU: Extracted motion metrics and synthetic linear acceleration / angular velocity signals.
• Natural Language Semantic Annotations: Continuous descriptive actions describing step-by-step driver events.
• Structured Export formats: Readily importable HDF5 structure and flattened CSV format for traditional ML models.

---

Dataset Structure

The dataset contains the main metadata, annotations, videos, and raw frame exports organized under a single structure.

Folder Layout

car-driving-multimodal-v1/
│
├── README.md
├── LICENSE
├── .gitattributes
│
└── Car_Driving/
    ├── rgb.mp4                          # Original egocentric RGB video
    ├── depth.mp4                        # Estimated monocular depth video
    ├── visualization_skeleton.mp4       # Video visualizing the estimated hand skeletons
    ├── metadata.json                    # Overall video and frame processing metadata
    ├── manifest.json                    # Full execution manifest with timestamp indices
    ├── hand_keypoints_2d.json           # Frame-wise pixel coordinates for 2D hand joints
    ├── hand_keypoints_3d.json           # Frame-wise metric 3D hand joint locations
    ├── hand_object_interactions.json    # Hand-to-object contact annotations per frame
    ├── object_annotations.json          # Bounding boxes and labels of detected objects
    ├── actions.json                     # Ground-truth action label categories per frame
    ├── semantic_actions.json            # Frame-aligned natural language action descriptions
    ├── motion_statistics.json           # Dynamic metrics (velocity, acceleration, jerk)
    ├── pseudo_imu.json                  # Synthesized 6-DOF IMU (accelerometer, gyroscope)
    ├── trajectories.json                # Wrist and finger-tip motion trajectories
    ├── task_summary.json                # Execution metadata and summary statistics
    ├── depth_metadata.json              # Technical properties of the depth camera/estimation
    ├── combined_dataset.csv             # Flattened tabular representations of all modalities
    ├── dataset.h5                       # Structured hierarchical HDF5 binary package
    ├── summary_report.txt               # Plain-text execution report
    ├── task_summary.txt                 # Plain-text task summary
    └── rgb_frames/                      # Directory containing 100 extracted JPEG frames (000001.jpg - 000100.jpg)

Annotation Files

File	Description
metadata.json	General video configuration and capture characteristics
hand_keypoints_2d.json	Frame-wise 2D coordinates for 21 hand landmarks
hand_keypoints_3d.json	Frame-wise 3D spatial points relative to the wrist
hand_object_interactions.json	Contact events mapping hands to active objects
object_annotations.json	Extracted object categories and bounding boxes
actions.json	Action activity categorization tags
semantic_actions.json	Continuous natural language activity descriptions
motion_statistics.json	Computed velocity, acceleration, and jerk profiles
pseudo_imu.json	Synthetic linear acceleration and angular rates
trajectories.json	Spatial paths of major joints across frames
combined_dataset.csv	Unified tabular annotations mapping all metrics per frame
dataset.h5	Hierarchical binary format storing the entire multimodal dataset
rgb.mp4	Egocentric driving camera feed
depth.mp4	Visual monocular depth estimation
visualization_skeleton.mp4	Overlay of 3D skeletal tracks onto the original RGB video

---

Dataset Statistics

The following statistics describe the version 1 release:

• Total Frames: 100 frames
• Video Duration: 30.07 seconds
• FPS: 29.97 FPS
• Resolution: 1024 x 576 pixels
• Detected Object Occurrences: 194 detections (Unique labels: person, car)
• Hand-Object Interactions: 14 active interactions
• Annotated Actions/Activities: 100 steps (Unique categories: pick, type_or_write, reach, moving, idle)

---

Usage

Below are examples of how to load and parse this dataset in Python.

1. Reading metadata.json

import json

with open("Car_Driving/metadata.json", "r") as f:
    metadata = json.load(f)

print(f"Video Resolution: {metadata['video_info']['width']}x{metadata['video_info']['height']}")
print(f"FPS: {metadata['video_info']['fps']:.2f}")

2. Loading dataset.h5

import h5py

with h5py.File("Car_Driving/dataset.h5", "r") as hf:
    # Print the available datasets in the H5 hierarchy
    print("Keys in H5:", list(hf.keys()))
    
    # Extract 3D hand keypoints and actions
    keypoints_3d = hf["hand_keypoints_3d"][:]
    actions = hf["actions"][:]
    
    print(f"Loaded 3D Keypoints shape: {keypoints_3d.shape}")

3. Loading combined_dataset.csv

import pandas as pd

df = pd.read_csv("Car_Driving/combined_dataset.csv")
print(df.head())

4. Parsing hand_keypoints_3d.json

import json

with open("Car_Driving/hand_keypoints_3d.json", "r") as f:
    keypoints_data = json.load(f)

# Print keypoints for the first frame
first_frame_keypoints = keypoints_data[0]["keypoints"]
print("First frame wrist position:", first_frame_keypoints[0]) # Wrist landmark (index 0)

---

Applications

• Egocentric Cabin Vision: Recognizing steering, infotainment adjustments, and phone interactions.
• Robotic Control & Imitation: Modeling human dexterity and hand postures for robotic manipulation.
• Hand Pose Estimation Benchmarking: 2D-to-3D pose mapping from monocular ego-views.
• Action Recognition: Fine-grained activity classification and temporal action localization (TAL).
• Vision-Language-Action (VLA) Models: Correlating natural language action labels with real physical hand dynamics.

---

Citation

If you use this dataset in your research, please cite:

@misc{cardrivingmultimodalv1_2026,
  title={Car Driving Multimodal Dataset v1},
  author={DeepAnnotate AI},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/deepannotateai/car-driving-multimodal-v1}},
  note={Version 1.0}
}

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Version History

• v1.0 (July 2026): Initial public release including video sequences, 2D/3D hand pose estimation, and synthetic sensory mappings.

Acknowledgements

This dataset was generated using the Hand Egocentric Multimodal Annotation Pipeline developed for this project.