Data Collection API

CARLA-based multi-modal sensor data pipeline for the LANCER project carla 0.9.14+ Python 3.8+

The carla-data-api module automates the collection of synchronized sensor data from a simulated urban environment. A Dodge Charger 2020 ego vehicle (EV) drives a fixed route while RGB, depth, semantic segmentation, and LiDAR sensors capture the scene simultaneously from three horizontal angles. The process is repeated automatically for every weather preset in WEATHER_PRESETS.csv, producing a structured dataset suitable for training perception and navigation models.

The pipeline iterates over all defined weather conditions, resetting the world between each run so that traffic and lighting conditions change consistently across the dataset.

Output Format

Each captured frame is saved as a single 960 × 960 px PNG containing a 4-row × 3-column sensor grid. Rows correspond to sensor modalities; columns correspond to the three horizontal viewing angles (left −120°, forward 0°, right +120°). Each cell is 240 × 320 px.

Grid Layout

The LiDAR row is produced by projecting the single roof-mounted point cloud onto each of the three camera planes using a pinhole camera model. Point brightness encodes distance logarithmically: bright pixels are close, dark pixels are far.

Frames are written to:

dataset/
└── <weather_preset_name>/
    ├── frame_0.png
    ├── frame_1.png
    └── ...

Sample Frames

The following samples were captured in clearday and clearnight conditions from the sample_dataset/ directory.

Clear Day · frame_0

Clear Day · frame_1

Clear Night · frame_0

Clear Night · frame_1

Requirements

Install the Python dependencies with:

pip install numpy opencv-python scikit-image

Usage

Quick Start

1. Start the CARLA server

   Launch the CARLA simulator. The server must be listening on localhost:2000 (the default port) before running the script.

   # Linux
   ./CarlaUE4.sh
   
   # Windows
   CarlaUE4.exe

2. Navigate to the API directory

   cd carla-data-api

3. Run the collection script

   python collect_data.py

   This executes generate_dataset(max_images_per_weather=2, interval=20), collecting 2 frames per weather preset with 20 simulation ticks (~2 s) between captures, across the first 6 weather presets in WEATHER_PRESETS.csv. Output is written to dataset/.

Custom Collection Parameters

Call generate_dataset with different arguments to control the volume and timing of captured data:

from collect_data import generate_dataset

generate_dataset(
    max_images_per_weather=10,   # frames to capture per weather condition
    interval=10                  # simulation ticks between captures (~1 s each)
)

Single-Weather Run

For finer control, call run_simulation directly to collect data for one weather condition without cycling through the full preset list:

import carla
from collect_data import Simulation, run_simulation

client = carla.Client('localhost', 2000)
client.set_timeout(30.0)
sim = Simulation(client)

run_simulation(
    sim=sim,
    weather={
        'sun_altitude_angle': 60.0,
        'cloudiness': 20.0,
        # ... other WeatherParameters fields
    },
    max_images=5,
    interval=15,
    save_path_prefix='my_dataset/clearday/frame'
)

Configuration

Weather Presets (WEATHER_PRESETS.csv)

Each row defines a named weather condition. The name column is used as the output subdirectory name. All other columns map directly to carla.WeatherParameters attributes.

Vehicle Probabilities (VEHICLE_PROBABILITIES.csv)

Controls the NPC traffic composition. The global category mix is fixed in the code; the CSV controls how individual models are distributed within each category.

Global category spawn mix (defined in Simulation.__init__):

EV Route

The ego vehicle follows a fixed sequence of map spawn-point indices defined in Simulation.get_spawn_points_and_drive_indices(). To change the route, edit the drive_indices list in that method. Collection stops early for a given weather run if the EV comes within 5 m of the penultimate route waypoint.

Sensor Setup

All sensors are mounted on the EV roof at a height of 2 × bounding_box.extent.z + 1 m. The LiDAR sensor faces forward; the three camera sets share the same mount point but are rotated to their respective yaw angles.

Collection Pipeline

The script runs in synchronous mode at a fixed 10 Hz (0.1 s/tick). The Traffic Manager is bound to port 8000.

1. WeatherIterator loads all presets from WEATHER_PRESETS.csv and yields them one by one.
2. For each preset, run_simulation reloads the CARLA world, applies the weather parameters, and spawns NPC traffic (up to 200 pedestrians and a configurable number of vehicles sampled from VEHICLE_PROBABILITIES.csv).
3. The ego vehicle is spawned, teleported to the route start, and handed to the Traffic Manager autopilot with lane-changing disabled.
4. Sensors are attached and their callbacks push raw data into a thread-safe Queue.
5. collect_data ticks the simulation, drains the queue after each tick, and assembles the 4-row × 3-column composite image every interval ticks.
6. Frames are saved to dataset/<preset_name>/frame_N.png until max_images_per_weather is reached or the EV nears the end of the route.
7. All sensors and NPC actors are destroyed before the next weather preset begins.

Notes