System and method for free space estimation
AssigneeDEKA Products Limited Partnership
InventorsRaajitha Gummadi, Christopher C. Langenfeld, Michael J. Slate, Christopher J. Principe
Life of this patent
Estimating the term…
Abstract
A system and method for estimating free space including applying a machine learning model to camera images of a navigation area, where the navigation area is broken into cells, synchronizing point cloud data from the navigation area with the processed camera images, and associating probabilities that the cell is occupied and object classifications of objects that could occupy the cells with cells in the navigation area based on sensor data, sensor noise, and the machine learning model.
Claims32
A method for estimating free space based on image data and point cloud data, the free space used for navigating an autonomous vehicle, the method comprising:
- receiving the image data and the point cloud data into the autonomous vehicle;
- semantically classifying the image data based on a machine learning model forming point classifications and point classification probabilities;
- associating each point in the point cloud data to the image data that are spatially co-located with the point cloud data;
- performing a first transform on the points in the point cloud data into an image coordinate system associated with the image data;
- classifying each of the first transformed points that represents an obstructed space and a non-obstructed space based on the spatial association of the first transformed points with the semantically classified image data forming obstructed points and non-obstructed points;
- performing a second transform on the first transformed points into a robot coordinate system associated with the autonomous vehicle;
- classifying each of the second transformed points that represents the non-obstructed space and the obstructed space within a pre-selected area surrounding the autonomous vehicle, forming a grid of the obstructed points and non-obstructed points, based on spatial association of the first transformed points with the semantically classified image data having point classifications and point probabilities;
- associating the obstructed points with a first probability based at least on (a) noise in the point cloud data, (b) a second probability that the point cloud data are reliable, and (c) a third probability that the point classifications are correct;
- and
- estimating the free space in the pre-selected area by computing a fourth probability based at least on (1) noise in the point cloud data, (2) the second probability, (3) a distance from the non-obstructed points to the obstructed space closest to the non-obstructed points, (4) the third probability, and (5) presence of the non-obstructed space.
A system for estimating free space based on image data and point cloud data, the free space used for navigating an autonomous vehicle, the autonomous vehicle having a front, a left side, and a right side, the system comprising:
- a pre-processor receiving camera image data from at least one camera into the autonomous vehicle, the pre-processor semantically classifying each pixel of the image data into a classification and calculating a probability associated with the classification, the classification and the probability being determined by a machine learning model;
- and
- a free space estimator including:
- a 3D point cloud to 3D image processor transforming the 3D point cloud to 3D image coordinates;
- a 3D image to 2D RGB transform transforming the 3D image coordinates to 2D RGB coordinates;
- a 2D to 3D baselink transform transforming the 2D RGB coordinates to 3D baselink coordinates forming transformed point cloud data;
- and
- a layer processor computing an obstacle classification layer, a probability layer, and a log odds layer based on the transformed point cloud data and the image data.
A method for mapping point cloud data from at least one LIDAR sensor onto semantically segmented image data from a plurality of cameras, the at least one LIDAR sensor and the plurality of cameras being located upon an autonomous vehicle, the method comprising:
- accessing a synchronizer, the synchronizer providing time synchronized point cloud data and the semantically segmented image data that are based on the time stamps on point cloud data and the semantically segmented image data that are received simultaneously, forming time-synchronized point cloud data from the time stamped point cloud data and time-synchronized semantically segmented image data from the time stamped semantically segmented image data;
- receiving time-synchronized point cloud data and the time-synchronized semantically segmented image data from the plurality of cameras, the plurality of cameras being mounted on a front of the autonomous vehicle, on a left side of the autonomous vehicle, and on an opposing right side of the autonomous vehicle;
- translating the time-synchronized point cloud data from a LIDAR coordinate system associated with the at least one LIDAR sensor to an image coordinate system associated with the plurality of cameras including:
- creating a rotation by applying a roll of −90° and a yaw of −90° to the point cloud data, the rotation producing rotated point cloud data in a 3D frame of reference according to a rotation matrix R;
- applying camera rotation matrices r to the rotated point cloud data to align a LIDAR frame of reference of the at least one LIDAR sensor with a camera frame of reference associated with the plurality of cameras;
- converting LIDAR points associated with the aligned LIDAR frame of reference to 3D image points by applying the rotation matrix R and camera translation factors tx/ty/tz to the rotated point cloud data;
- accessing a calibration matrix K associated with the plurality of cameras;
- applying the calibration matrix K to the 3D image points forming calibrated 3D image points (x,y,z);
- and
- converting the calibrated 3D image points (x,y,z) to 2D points (Xrgb,Yrgb) wherein Xrgb=x/z, Yrgb=y/z;
- and
- providing each of the 2D points (Xrgb,Yrgb) a depth value of a LIDAR point that is closest to the autonomous vehicle.
Dependencies are inferred from the claim text.