This is an experimental implementation of a runtime monitor for self-driving cars that verifies aspects of perception in both LiDAR and vision. In each example, a controller passes a certificate to the monitor, which uses the certificate to either confirm or deny the safety of the situation.
To run a script inside the Carla directory, do the following from the top level interlock/ directory: python3 -m carla_scripts.<script_name> Note the missing .py. This allows scripts in the Carla directory to import the checkers without having to edit the path variable.
The controller filters snow particles from the LiDAR point cloud and generates a certificate containing points that are far enough from the source to ensure no collisions will happen in the near future.
The monitor checks that the LiDAR points in the certificate satisfy two conditions: sufficient spread and sufficient density. To achieve sufficient spread, the certificate must span the size of one lane in front of the car, both vertically and horizontally. To achieve sufficient density, the certificate must contain points from enough LiDAR scan rows, and must have no horizontal gaps of a size greater than a pre-specified parameter.
The controller takes pixels from the camera image's lane lines and computes the LiDAR points that correspond to them. It generates a certificate containing those points and the ground plane (computed from the LiDAR point cloud).
The monitor checks that the LiDAR points corresponding to lane line points are on the ground plane.
The controller generates a certificate containing a proposed pair of lane lines (given as polynomials in a top-down view), a transformation matrix for converting from top-down to camera view, an image of the road, and a set of color filtering thresholds appropriate for the current lighting conditions
The monitor checks that:
- The lane lines from the polynomials are adequately spaced apart and parallel
- The polynomials match closely with the image (using the transformation matrix to get the corresponding points in the camera view and the color filtering thresholds for processing the image)
The controller generates a certificate to validate its segmentation. We implement the following checks for the segmentation certificate:
- Density: there must be a sufficient coverage of the LiDAR field of view
- Contiguity: LiDAR points making up a classified object must have similar velocity and must be close enough to one another
- Ground Plane: LiDAR points making up a ground plane must be of a sufficient height
- Collision Freedom: no object can pose an immediate collision risk