YuanYunshuang/cosense-simulation

Simulation for collective perception with CALAR & SUMO

Run the simulation

Generate routes and rerouting strategy

• run python script trafficgen.py under directory trafic_flow to regenerate routes if you made some modifications on road generation. Otherwise routes are already generated in folder town5.

Activate Carla simulation server

Cofigurate simulation parameters and start simulation

• config parameters are in config.json file under dir Sumo, path for saving data is also defined there.
  • "root_path": where to save the generated data
  • "sensor_names": which sensors you want to use for collecting data
• run the simulation under the project root dir

python main.py traffic_flow/town5/Town05.sumocfg --tls-manager carla --sumo-gui

Simulation raw data

Data structure

• rootdir
  • vehicle_id...
    • camera
      • image...
      • image_meta...
    • camera_sem
      • images_sem_label...
      • image_sem_meta...
    • lidar
      • pointcloud...
      • pointcloud_meta...
    • lidar_sem
      • pointclouds_sem_label...
      • pointclouds_sem_meta...
  • info.csv

Definition of file names and file contents

• ... means there are multiple such file or folders
• rootdir: j xxx e xxx states for junction, the junction number, ego and ego vehicle id respectively
• vehicle_id: vehicle_id
• image, image_sem_label: RGB image data in png format, file name = ".png", check carla documentation for color definition
• image_meta, image_sem_meta: meta information about the image with the same frame name, file name = "_meta.txt"
  • line1: frame, timestamp, fov, height, width
  • line2: roll, pitch, yaw, x, y, z
• pointcloud, pointclouds_sem_label: pointcloud raw data, file name = ".pcd", check carla documentation for color definition
• pointcloud_meta, pointclouds_sem_meta: meta information about the pointcloud with the same frame name, file name = "_meta.txt"
  • line1: frame, timestamp, horizontal_angle, n_channels
  • line2: roll, pitch, yaw, x, y, z
  • line3: 64 elements array, each element indicate the points number measured by the corresponding laser.

Data formating

• run the python script formatting_data.py in folder scripts/python , before runing the script, you need to change the path(in_path) to the generated raw simulation data as well as the path(out_path) where you want to store the formatted data.

Data structure

• the formatted data has 3 folders

  • cloud_ego: clouds collected by ego vehicle, no down-sampling
  • cloud_coop: point clouds of cooperative vehicles
  • cloud_fused: fused clouds(including ego-vehicle cloud), down-sampled through voxel grid(size=0.2m)
  • label_box: ground truth bounding boxes of all vehicles in the scene of each frame

• point cloud files:

  • all data are written in binary files. Data are all in float32 length,
  • each data point has 4 columns(x,y,z,label), following is the semantic meaning of labels

      LABEL_COLORS = np.array([
      (0, 0, 0),       #0 Unlabeled
      (70, 70, 70),    #1 Buildings
      (100, 40, 40),   #2 Fences
      (55, 90, 80),    #3 Other
      (220, 20, 60),   #4 Pedestrians
      (153, 153, 153), #5 Poles
      (157, 234, 50),  #6 RoadLines
      (128, 64, 128),  #7 Roads
      (244, 35, 232),  #8 Sidewalks
      (107, 142, 35),  #9 Vegetation
      (0, 0, 142),     #10 Vehicles
      (102, 102, 156), #11 Walls
      (220, 220, 0),   #12 TrafficSign
      (70, 130, 180),  #13 Sky
      (81, 0, 81),     #14 Ground
      (150, 100, 100), #15 Bridge
      (230, 150, 140), #16 Railtrack
      (180, 165, 180), #17 GuardRail
      (250, 170, 30),  #18 TrafficLight
      (110, 190, 160), #19 Static
      (170, 120, 50),  #20 Dynamic
      (45, 60, 150),   #21 Water
      (145, 170, 100)  #22 Terrain
      ])

• filename: <junction_id>_<frame_id>.<extention>

• label_box format:

  • each row indicates one bounding box,
  • column from left to right are: vehicle_id, vehicle_class, x, y, z, rx, ry, rz, l, w, h
  • x,y,z: the center point of the bounding box
  • rx, ry, rz: orientation
  • l, w, h: vehicle size