sebnem-byte/drone-navigation-event-camera

Edged-USLAM Dataset: Drone Navigation with Event Camera

🔗 Citation

@inproceedings{sariozkan2026edged,
  title={Edged USLAM: Edge-Aware Event-Based SLAM with Learning-Based Depth Priors},
  author={Sarıözkan, Şebnem and Şahin, Hürkan and Álvarez-Tuñón, Olaya and Kayacan, Erdal},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}

ℹ️ Extra Info

Project Page

This dataset contains synchronized Event Camera (DAVIS346), IMU, and Ground Truth (Motion Capture) data recorded for the evaluation of Edged-USLAM The recordings include diverse motion profiles (aggressive 6-DoF, square, line trajectories) and challenging illumination conditions (Low-light, HDR, Dynamic Lighting).

📂 2. Dataset Structure

The dataset is organized into three main folders. Please pay attention to the Ground Truth topic as it changes depending on the folder.

**1. motion/ (Motion Categories) **

• Description: Aggressive and distinct trajectory maneuvers.
• Sequences:
  • line.bag
  • square.bag
  • turner.bag (Aggressive turning)
  • (Other motion sequences like circle, aggressive are included here)
• 📍 Ground Truth Topic: /mavros/vision_pose/pose

filtered/

• Description: Recorded with an IR-filtered lens attached to the DAVIS346.
• Effect: Improves robustness under strong illumination but reduces sensitivity.
• Sequences: Includes illumination variations (low-lit, dynamic HDR, constant HDR, 30% lit, 60% lit).
• 📍 Ground Truth Topic: /local_pose_vicon/pose

unfiltered/

• Description: Recorded with the Default (non-filtered) lens.
• Effect: Preserves the original DAVIS346 response, including IR wavelengths.
• Sequences: Includes illumination variations (low-lit, dynamic HDR, constant HDR, etc.) for spectral comparison.
• 📍 Ground Truth Topic: /local_pose_vicon/pose

---

**2. Illumination/ (Illumination Category) **

These folders contain identical flight paths recorded with different lens configurations to test photometric robustness.

• Focus: Robustness against HDR, Low-light, and sudden lighting changes.
• Sequences included (in both folders):
  • low-lit (< 5 Lux)
  • 30% lit
  • 60% lit
  • dynamic HDR (Blinking lights)
  • constant HDR (Strong side light/Sunlight)
• Folder Differences:
  • filtered/: IR-filtered lens (Better for HDR, less sensitive in dark).
  • unfiltered/: Default lens (High sensitivity, includes IR spectrum).
• 📍 Ground Truth Topic: /local_pose_vicon/pose

---

📝 3. Notes

• Synchronization: All sequences are timestamp-synchronized across topics.
• Ground Truth Frame: In both cases, the pose provides 6-DoF UAV pose in the Vicon frame.
• Performance: As shown in our paper (Table III), motion sequences test geometric robustness, while filtered/unfiltered sequences test photometric robustness under HDR and low-light.

---

📡 1. Sensor Topics

Topic	Type	Description
/dvs/events	Event Stream	DAVIS346 asynchronous events
/dvs/image_raw	Image	Grayscale intensity frames (APS)
/dvs/imu	IMU	Accelerometer + Gyroscope (200Hz)
/mavros/vision_pose/pose	Pose	GT for motion/ folder
/local_pose_vicon/pose	Pose	GT for filtered/ & unfiltered/ folders

---

⚙️ Calibration Parameters (DAVIS-IJRR17)

Camera Model: Pinhole with Radial-Tangential Distortion
Resolution: 346 x 260

# Intrinsics
fx: 254.6368
fy: 253.6548
cx: 168.9970
cy: 121.0141

# Distortion (Radial-Tangential)
k1: -0.3755
k2: 0.1311
p1: 0.00137
p2: -0.00159

# Extrinsics (Camera to IMU/Body)
# Translation (m): [0.0067, 0.0007, 0.0343] (approx 3.4cm offset in Z)
# Rotation: Identity matrix