Method for generating at least one ground truth from a bird's eye view
A method for generating at least one image from a bird's eye view. The method includes: a) carrying out a sensor data point cloud compression; b) carrying out a point cloud filtering in a camera perspective; c) carrying out an object completion; and d) carrying out a bird's eye view segmentation and generating an elevation map.
1 . A method for generating at least one representation from a bird's eye view, the method comprising the following steps:
a) carrying out a sensor data point cloud compression;
b) carrying out a point cloud filtering in a perspective of a camera;
c) carrying out an object completion; and
d) carrying out a bird's eye view segmentation and generating an elevation map;
wherein the method includes at least one of the following features (I)-(III):
(I) the cloud compression includes at least one of:
(i) compressing data points from respective frames of a plurality of respective points in time into a single frame of data points; and
(ii) removing respective data points based on the respective data points corresponding to dynamic objects;
(II) the point cloud filtering includes at least one of:
(i) removing data points that correspond to a location that is outside a viewing frustum of the camera; and
(ii) applying the data points to a matrix having extrinsic parameters of pose of the camera in space and intrinsic parameters of the camera; and
(III) the object completion includes at least one of:
(i) a neural network completing partially visible objects to which at least a part of the point cloud corresponds; and
(ii) carrying out a morphological operation that fills in a missing portion of an incomplete object to which the at least the part of the point cloud corresponds.
2 . The method according to claim 1 , wherein the representation is also at least based on sensor data obtained from at least one active surroundings sensor, wherein the at least one active surroundings sensor includes a LiDAR sensor and/or a radar sensor.
3 . The method according to claim 1 , wherein, in step b), at least one camera parameter is used to back-project sensor data points onto a current image area.
4 . The method according to claim 1 , wherein, in step c), at least one cuboid box is projected onto a current bird's eye view region.
5 . The method according to claim 1 , wherein in step d), all 3D points for a valid bird's eye view region are collected.
6 . The method according to claim 1 , wherein the cloud compression includes the compressing of the data points from the respective frames of the plurality of respective points in time into the single frame of data points.
7 . The method according to claim 1 , wherein the cloud compression includes the removing of the respective data points based on the respective data points corresponding to the dynamic objects.
8 . The method according to claim 7 , wherein, in step a), the removing is performed based on a splitting of a sensor data point cloud is split into static and dynamic object points on the basis of semantic information.
9 . The method according to claim 1 , wherein the point cloud filtering includes the removing of the data points that correspond to the location that is outside the viewing frustum of the camera.
10 . The method according to claim 1 , wherein the point cloud filtering includes the applying of the data points to the matrix having the extrinsic parameters of pose of the camera in space and the intrinsic parameters of the camera.
11 . The method according to claim 1 , wherein the object completion includes the neural network completing the partially visible objects to which the at least the part of the point cloud corresponds.
12 . The method according to claim 1 , wherein the object completion includes the carrying out of the morphological operation that fills in the missing portion of the incomplete object to which the at least the part of the point cloud corresponds.
13 . The method according to claim 12 , wherein, in step c), the morphological operation is applied to at least one object of a current bird's eye view region.
14 . The method according to claim 12 , wherein the morphological operation is a dilation.
15 . The method according to claim 1 , wherein:
the cloud compression includes the compressing of the data points from the respective frames of the plurality of respective points in time into the single frame of data points;
the respective points in time include a keyframe point in time and non-keyframe points in time that are at least one of before and after the keyframe point in time; and
the cloud compression further includes performing for each of the non-keyframe points in time, the removing of the respective data points based on the respective data points corresponding to the dynamic objects, while keeping respective data points corresponding to dynamic objects in the keyframe.
16 . A non-transitory machine-readable storage medium on which is stored a computer program for generating at least one representation from a bird's eye view, the computer program, when executed by a computer, causing the computer to perform a method having the following steps:
a) carrying out a sensor data point cloud compression;
b) carrying out a point cloud filtering in a perspective of a camera;
c) carrying out an object completion; and
d) carrying out a bird's eye view segmentation and generating an elevation map;
wherein the method includes at least one of the following features (I)-(III);
(I) the cloud compression includes at least one of:
(i) compressing data points from respective frames of a plurality of respective points in time into a single frame of data points; and
(ii) removing respective data points based on the respective data points corresponding to dynamic objects;
(II) the point cloud filtering includes at least one of:
(i) removing data points that correspond to a location that is outside a viewing frustum of the camera; and
(ii) applying the data points to a matrix having extrinsic parameters of pose of the camera in space and intrinsic parameters of the camera; and
(III) the object completion includes at least one of:
(i) a neural network completing partially visible objects to which at least a part of the point cloud corresponds; and
(ii) carrying out a morphological operation that fills in a missing portion of an incomplete object to which the at least the part of the point cloud corresponds.
17 . An object recognition system for a vehicle, the system configured to generate at least one representation from a bird's eye view, the system configured to
a) carry out a sensor data point cloud compression;
b) carry out a point cloud filtering in a perspective of a camera;
c) carry out an object completion; and
d) carry out a bird's eye view segmentation and generate an elevation map;
wherein the object recognition system includes at least one of the following features (I)-(III):
(I) the cloud compression includes at least one of:
(i) compressing data points from respective frames of a plurality of respective points in time into a single frame of data points; and
(ii) removing respective data points based on the respective data points corresponding to dynamic objects;
(II) the point cloud filtering includes at least one of:
(i) removing data points that correspond to a location that is outside a viewing frustum of the camera; and
(ii) applying the data points to a matrix having extrinsic parameters of pose of the camera in space and intrinsic parameters of the camera; and
(III) the object completion includes at least one of:
(i) a neural network completing partially visible objects to which at least a part of the point cloud corresponds; and
(ii) carrying out a morphological operation that fills in a missing portion of an incomplete object to which the at least the part of the point cloud corresponds.