Sensors and data

Sensors are actors that retrieve data from their surroundings. They are crucial to create learning environment for driving agents.

This page summarizes everything necessary to start handling sensors. It introduces the types available and a step-by-step guide of their life cycle. The specifics for every sensor can be found in the sensors reference.


Sensors step-by-step

The class carla.Sensor defines a special type of actor able to measure and stream data.

  • What is this data? It varies a lot depending on the type of sensor. All the types of data are inherited from the general carla.SensorData.
  • When do they retrieve the data? Either on every simulation step or when a certain event is registered. Depends on the type of sensor.
  • How do they retrieve the data? Every sensor has a listen() method to receive and manage the data.

Despite their differences, all the sensors are used in a similar way.

Setting

As with every other actor, find the blueprint and set specific attributes. This is essential when handling sensors. Their attributes will determine the results obtained. These are detailed in the sensors reference.

The following example sets a dashboard HD camera.

# Find the blueprint of the sensor.
blueprint = world.get_blueprint_library().find('sensor.camera.rgb')
# Modify the attributes of the blueprint to set image resolution and field of view.
blueprint.set_attribute('image_size_x', '1920')
blueprint.set_attribute('image_size_y', '1080')
blueprint.set_attribute('fov', '110')
# Set the time in seconds between sensor captures
blueprint.set_attribute('sensor_tick', '1.0')

Spawning

attachment_to and attachment_type, are crucial. Sensors should be attached to a parent actor, usually a vehicle, to follow it around and gather the information. The attachment type will determine how its position is updated regarding said vehicle.

  • Rigid attachment. Movement is strict regarding its parent location. This is the proper attachment to retrieve data from the simulation.
  • SpringArm attachment. Movement is eased with little accelerations and decelerations. This attachment is only recommended to record videos from the simulation. The movement is smooth and "hops" are avoided when updating the cameras' positions.
  • SpringArmGhost attachment. Like the previous one but without doing the collision test, so the camera or sensor could cross walls or other geometries.
transform = carla.Transform(carla.Location(x=0.8, z=1.7))
sensor = world.spawn_actor(blueprint, transform, attach_to=my_vehicle)

Important

When spawning with attachment, location must be relative to the parent actor.

Listening

Every sensor has a listen() method. This is called every time the sensor retrieves data.

The argument callback is a lambda function. It describes what should the sensor do when data is retrieved. This must have the data retrieved as an argument.

# do_something() will be called each time a new image is generated by the camera.
sensor.listen(lambda data: do_something(data))

...

# This collision sensor would print everytime a collision is detected. 
def callback(event):
    for actor_id in event:
        vehicle = world_ref().get_actor(actor_id)
        print('Vehicle too close: %s' % vehicle.type_id)

sensor02.listen(callback)

Data

Most sensor data objects have a function to save the information to disk. This will allow it to be used in other environments.

Sensor data differs a lot between sensor types. Take a look at the sensors reference to get a detailed explanation. However, all of them are always tagged with some basic information.

Sensor data attribute Type Description
frame int Frame number when the measurement took place.
timestamp double Timestamp of the measurement in simulation seconds since the beginning of the episode.
transform carla.Transform World reference of the sensor at the time of the measurement.


Important

is_listening is a sensor attribute that enables/disables data listening at will.
sensor_tick is a blueprint attribute that sets the simulation time between data received.


Types of sensors

Cameras

Take a shot of the world from their point of view. For cameras that return carla.Image, you can use the helper class carla.ColorConverter to modify the image to represent different information.

  • Retrieve data every simulation step.
Sensor Output Overview
Depth carla.Image Renders the depth of the elements in the field of view in a gray-scale map.
RGB carla.Image Provides clear vision of the surroundings. Looks like a normal photo of the scene.
Optical Flow carla.Image Renders the motion of every pixel from the camera.
Semantic segmentation carla.Image Renders elements in the field of view with a specific color according to their tags.
Instance segmentation carla.Image Renders elements in the field of view with a specific color according to their tags and a unique object ID.
DVS carla.DVSEventArray Measures changes of brightness intensity asynchronously as an event stream.



Detectors

Retrieve data when the object they are attached to registers a specific event.

  • Retrieve data when triggered.
Sensor Output Overview
Collision carla.CollisionEvent Retrieves collisions between its parent and other actors.
Lane invasion carla.LaneInvasionEvent Registers when its parent crosses a lane marking.
Obstacle carla.ObstacleDetectionEvent Detects possible obstacles ahead of its parent.


Other

Different functionalities such as navigation, measurement of physical properties and 2D/3D point maps of the scene.

  • Retrieve data every simulation step.
Sensor Output Overview
GNSS carla.GNSSMeasurement Retrieves the geolocation of the sensor.
IMU carla.IMUMeasurement Comprises an accelerometer, a gyroscope, and a compass.
LIDAR carla.LidarMeasurement A rotating LIDAR. Generates a 4D point cloud with coordinates and intensity per point to model the surroundings.
Radar carla.RadarMeasurement 2D point map modelling elements in sight and their movement regarding the sensor.
RSS carla.RssResponse Modifies the controller applied to a vehicle according to safety checks. This sensor works in a different manner than the rest, and there is specific RSS documentation for it.
Semantic LIDAR carla.SemanticLidarMeasurement A rotating LIDAR. Generates a 3D point cloud with extra information regarding instance and semantic segmentation.



That is a wrap on sensors and how do these retrieve simulation data.

Thus concludes the introduction to CARLA. However there is yet a lot to learn.

  • Continue learning. There are some advanced features in CARLA: rendering options, traffic manager, the recorder, and some more. This is a great moment to learn more about them.
  • Experiment freely. Take a look at the References section of this documentation. It contains detailed information on the classes in the Python API, sensors, code snippets and much more.
  • Give your two cents. Any doubts, suggestions and ideas are welcome in the forum.