tudelft-iv/view-of-delft-dataset

What it solves

This project provides a comprehensive automotive dataset designed for training and evaluating machine learning models for road user detection and tracking in complex urban traffic. It addresses the need for high-quality, synchronized multi-sensor data with precise 3D bounding box annotations for various road users like pedestrians, cyclists, and cars.

How it works

The dataset consists of 8,600 frames of synchronized and calibrated data from four primary sources:

• LiDAR: A 64-layer Velodyne HDL-64 S3 scanner.
• Cameras: A stereo camera system.
• Radar: A ZF FRGen21 3+1D radar.
• Odometry: A filtered combination of RTK GPS, IMU, and wheel odometry.

These sensors are jointly calibrated to provide a spatial and temporal alignment of the environment. The project includes a development kit to help researchers load, transform, and visualize the data (2D/3D) and evaluate their models.

Who it’s for

Academic and non-profit research institutions, specifically Master's and PhD students and staff, for non-commercial research purposes.

Highlights

• Multi-modal data: Combines LiDAR, stereo camera, and 3+1D radar data.
• Extensive annotations: Over 123,000 3D bounding box labels for 13 road user classes, including pedestrians, cyclists, and cars.
• Detailed labels: Annotations include occlusion, activity, and tracking IDs to follow objects across frames.
• Development kit: Includes Jupyter Notebooks for data loading and visualization.