Data Generation

CARLA

• CarlaUtils.py

  • Utility functions for generating raw data from CARLA.
  • Generate autonomous pedestrian and vehicular traffic.
  • Callbacks for sensors to save data.
  • Create sensors from blueprints.

• GatherCarlaData.py

  • Obtain and record raw sensor data. Imports utilities from CarlaUtils.py.
  • Parameters are passed to the main function from a dictionary in the bottom of the file.
    • storage_dir (where output are stored)
    • new_world (town number)
    • num_vehicles (total vehicles)
    • num_walkers (total pedestrians)
    • num_sensors (total lidars)
  • To use, first start an instance of CARLA (i.e. ./CarlaUE4.sh), then run GatherCarlaData.py to connect and record data.
  • The directory generated is explained below
    • bev: bird-eye-view images
    • instancesXX: per point instance from the XX lidar
    • labelsXX: per point label from the XX lidar
    • poseXX: XX lidar position in the world
    • velocitiesXX: per point velocity from the XX lidar
    • velodyneXX: point cloud from the XX lidar
    • timesXX.txt: timestamp at each frame from the XX lidar
    • recording.log: log file that enables playback using Carla
    • More information about the sensor we use could be found in the official Carla wiki

Scene Generation

• SceneCompletion.py

  • Process raw CARLA data and generate semantic voxel grids.
  • Parameters are located in the main function at the bottom of the file.
    • cylindrical (boolean for storing voxels in cylindrical or cartesian coordinates)
    • parent_dir (where output are stored)
    • t_start, t_end (starting frame and ending frame for processing)
    • free_res (distance in meters to ray trace free space)
  • After raw data is generated, run this script to process the data and create a semantic volume.
  • A JSON file with the parameters of the semantic volume including size, resolution, and coordinate system, are saved in the folder "evaluations".

• ShapeContainer.py

  • Helper class for semantic volumes.

• DBKIData.py

  • Generate data similar to SceneCompletion.py, however in a format for running with semantic mapping algorithms.