Class labeling system for autonomous driving

A class labeling system for autonomous driving, comprising:

• a detection module configured to detect objects for autonomous driving from an image captured by a camera to generate a bounding box for each of the objects and to detect property information about each of the objects;
• a segmentation module configured to determine classes for each pixel of the bounding boxes detected by the detection module and to process at least one of the classes as don't care;
• and
• a lane road boundary detection module configured to detect at least one of lane and road boundaries using the bounding boxes detected by the detection module,
• wherein the lane road boundary detection module comprises:
• a free instance segmentation unit configured to output a pixel-wise classification and an instance of the image captured by the camera using an instance deep learning network;
• and
• a B-spline detection unit configured to generate a B-spline curve for lane or road shapes through the pixel-wise classification and the instance of the image that are output from the free instance segmentation unit.

An apparatus for autonomous driving, the apparatus comprising:

• at least one processor;
• and
• at least one memory comprising one or more sequences of one or more instructions that, in response to being executed by the at least one processor, cause the apparatus at least to:
• analyze an image to detect objects for autonomous driving;
• generate a bounding box for each of the objects, each of the bounding boxes comprising first pixels;
• crop each of the bounding boxes to detect second pixels from among the first pixels, the second pixels of a bounding box being related to a corresponding object in the bounding box;
• detect property information and a sub-property about each of the objects;
• determine classes for each of the first pixels;
• process at least one of the classes as being irrelevant with regard to the objects;
• receive light detection and ranging (LiDAR) distance input values;
• detect distances to the objects using the LiDAR distance input values;
• detect at least one of lane and road boundaries using the bounding boxes;
• detect at least one of a vehicle and a pedestrian using the bounding boxes;
• and
• predict a location of the at least one of the vehicle and the pedestrian.