UofTPed50 is a dataset that can be used for benchmarking the positional accuracy of 3D pedestrian detection. We provide accurate positioning information by attaching a GPS system to the pedestrian itself. This dataset consists of 50 sequences of varying distance, pedestrian trajectory, and ego-vehicle trajectory. Each sequence contains one pedestrian. The scenarios are broken into four groups:

1. 34 Sequences of a straight-line pedestrian trajectory with a stationary ego-vehicle at seven distances. (Seq. 1-34)

2. 3 sequences tracking straight lateral pedestrian trajectories with respect to a dynamic ego-vehicle. (Seq. 35-37)

3. 8 sequences tracking straight longitudinal trajectory trajectories, 4 with a static ego-vehicle and 4 with a dynamic ego-vehicle.

4. 6 sequences tracking complex trajectories (curves, zig-zag motion) with respect to a stationary ego-vehicle.

Data was collected on our self-driving car, Zeus, illustrated above. Sensor data includes a Velodyne HDL-64 3D LIDAR, a 5 MP monocular camera, and a Novatel PwrPak7 GPS/IMU with TerraStar corrections (< 10 cm reported position error). Position data for the pedestrian was collected by attaching a tethered antenna of a separate Novatel PwrPak7 GPS with TerraStar corrections.

To synchronize data between the ego-vehicle and the pedestrian, we use UTC timestamps.

You can download the current version of this dataset here

If this dataset was helpful in your work, please consider citing our CRV paper

Any questions or comments regarding the dataset can be sent to keenan (dot) burnett ‘at’ autodrive.utoronto.ca