Best practice for rear-view camera and left/right mirror cameras in CARLA 0.9.x

[Jeya0507]

Hello CARLA team and community,

we are developing a driving-school simulator based on CARLA 0.9.x and pygame.

For parking training, we need:

1. an automatic rear-view camera when reverse gear is selected,
2. a left exterior mirror,
3. a right exterior mirror.

Our current simulator already uses the standard CameraManager and one main cockpit camera.

Before implementing additional sensor.camera.rgb actors, we would like to ask:

• Is there an official CARLA example or recommended architecture for multiple simultaneous camera views?
• Does CARLA provide any built-in functional rear-view or side-mirror system?
• Should mirror images be produced by separate RGB cameras and horizontally flipped in the client UI?
• What attachment types are recommended for mirror cameras: Rigid or SpringArm?
• Are there known performance limitations when running three additional RGB cameras?
• Is there a recommended way to synchronize all camera images with the main camera?
• Are there example transforms for rear, left-mirror and right-mirror cameras on standard vehicles?
• Is there a recommended cleanup pattern when the vehicle is restarted or destroyed?

Our target vehicle is currently vehicle.nissan.patrol, and the UI is rendered with pygame.

Thank you for your guidance.

[yudin-s]

For this use case I would treat each mirror/rear view as its own sensor.camera.rgb actor attached to the vehicle. CARLA does not provide a built-in functional rear-view/mirror system that automatically renders vehicle mirrors for you.

The closest official examples are:

• PythonAPI/examples/visualize_multiple_sensors.py: renders several simultaneous sensors in one pygame window. It already uses four RGB cameras attached to one vehicle, with yaw values like -90, 0, 90, and 180.
• PythonAPI/examples/sensor_synchronization.py: shows the queue-based pattern for collecting one frame from each sensor in synchronous mode.
• PythonAPI/examples/manual_control.py: CameraManager is useful as a reference for camera blueprint setup and cleanup, but it is built around one active camera view at a time.

Recommended architecture:

• spawn one main cockpit camera plus separate RGB cameras for rear, left mirror, and right mirror
• attach them with carla.AttachmentType.Rigid for mirror/rear views, because you normally want stable vehicle-relative camera placement; SpringArm/SpringArmGhost is better for third-person chase cameras
• render each callback into its own pygame surface and blit those surfaces into the dashboard/mirror rectangles
• horizontally flip the mirror surfaces in the UI if you want real mirror semantics; CARLA camera output itself is just a camera view, not a mirror
• run synchronous mode if you need all views from the same simulation frame, and collect by image.frame as in sensor_synchronization.py
• destroy all spawned sensors on vehicle reset/restart before destroying the vehicle, or keep them in a list and call stop()/destroy() for each

Approximate starting transforms for vehicle.nissan.patrol should be tuned visually, but the pattern is:

rear_tf = carla.Transform(carla.Location(x=-2.5, z=1.6), carla.Rotation(yaw=180))
left_tf = carla.Transform(carla.Location(x=0.6, y=-1.1, z=1.35), carla.Rotation(yaw=-120))
right_tf = carla.Transform(carla.Location(x=0.6, y=1.1, z=1.35), carla.Rotation(yaw=120))

For performance, three extra RGB cameras can be expensive, especially at full window resolution. Give the mirror cameras smaller image_size_x / image_size_y, reduce fov to mirror-like values, and keep the sensor tick rate no higher than the UI actually needs.

If this gives you the implementation direction you needed, please mark it as the answer so other CARLA/pygame simulator users can find the multi-camera pattern.

[repa1030]

Hi Jeya0507,

I just stumbled across your question. Very interesting! We are currently working on an open simulator framework based on CARLA for driving, cycling and pedestrian simulation (https://arxiv.org/abs/2609.11478v1). Feel free to contact us; we can certainly learn a lot from each other! Do you also plan to open-source your framework?

Best regards
Patrick