Data Processing Method and Apparatus
Provided are a data processing method and apparatus. The method includes: After a plurality of ground truth boxes that correspond to a lidar point cloud and a millimeter-wave radar point cloud are obtained, for unification of coordinate systems of the plurality of ground truth boxes and the millimeter-wave radar point cloud, position transformation further needs to be performed on the plurality of ground truth boxes until in all the ground truth boxes, a proportion of a quantity of ground truth boxes whose quantity of millimeter-wave radar point clouds reaches a preset threshold in a total quantity of ground truth boxes reaches a preset proportion, and then the plurality of ground truth boxes on which position transformation is performed and the millimeter-wave radar point cloud are trained, to generate a target detection model. This can avoid an inaccurate training result caused because a reflection point exists at a scattering energy center and does not correspond to a position of a ground truth box due to a working principle of a millimeter-wave radar based on an electromagnetic wave, and can optimize a training dataset of the target detection model, to improve accuracy of the target detection model.
1 . A data processing method, comprising:
obtaining a lidar point cloud and a millimeter-wave radar point cloud;
obtaining a plurality of ground truth boxes corresponding to the lidar point cloud; and
performing position transformation on the plurality of ground truth boxes until a preset condition is met, wherein the preset condition is that a proportion of a quantity of ground truth boxes whose quantity of millimeter-wave radar point clouds reaches a preset threshold in the plurality of ground truth boxes reaches a preset proportion, the millimeter-wave radar point cloud and the plurality of ground truth boxes are located in a same coordinate system, and the plurality of ground truth boxes on which position transformation is performed and the millimeter-wave radar point cloud are a training dataset of a millimeter-wave detection model.
2 . The method according to claim 1 , wherein the position transformation comprises at least one of a translation operation, a rotation operation, and a scale-up operation, and a scaling factor of the scale-up operation is within a preset range.
3 . The method according to claim 1 or 2 , wherein after the performing position transformation on the plurality of ground truth boxes until a preset condition is met, the method further comprises:
clearing a ground truth box whose quantity of point clouds is less than the preset threshold in the plurality of ground truth boxes.
4 . The method according to any one of claims 1 to 3 , wherein the millimeter-wave radar point cloud comprises a target-level point cloud and/or an original point cloud, and before the performing position transformation on the plurality of ground truth boxes until a preset condition is met, the method further comprises:
performing motion compensation on the target-level point cloud when the millimeter-wave radar point cloud comprises the target-level point cloud.
5 . The method according to any one of claims 1 to 4 , wherein before the performing position transformation on the plurality of ground truth boxes until a preset condition is met, the method further comprises:
performing system delay compensation on the millimeter-wave radar point cloud.
6 . The method according to any one of claims 1 to 5 , wherein after the performing position transformation on the plurality of ground truth boxes until a preset condition is met, the method further comprises:
extracting a target feature from the millimeter-wave radar point cloud, wherein the target feature comprises a plurality of the following: coordinates, static/dynamic attributes, a radar cross section, an absolute velocity, a relative velocity, a radar type, and time sequence information of a superimposed frame; and
performing training by using the plurality of ground truth boxes on which position transformation is performed and the target feature as the training dataset.
7 . The method according to any one of claims 1 to 6 , wherein the millimeter-wave radar point cloud is generated by superimposing a plurality of frames of point clouds.
8 . A data processing apparatus, comprising:
an obtaining unit, configured to obtain a lidar point cloud and a millimeter-wave radar point cloud, and obtain a plurality of ground truth boxes corresponding to the lidar point cloud; and
a position transformation unit, configured to perform position transformation on the plurality of ground truth boxes until a preset condition is met, wherein the preset condition is that a proportion of a quantity of ground truth boxes whose quantity of millimeter-wave radar point clouds reaches a preset threshold in the plurality of ground truth boxes reaches a preset proportion, the millimeter-wave radar point cloud and the plurality of ground truth boxes are located in a same coordinate system, and the plurality of ground truth boxes on which position transformation is performed and the millimeter-wave radar point cloud are a training dataset of a millimeter-wave detection model.
9 . The apparatus according to claim 8 , wherein the position transformation comprises at least one of a translation operation, a rotation operation, and a scale-up operation, and a scaling factor of the scale-up operation is within a preset range.
10 . The apparatus according to claim 8 or 9 , wherein the apparatus further comprises a clearing unit, and the clearing unit is specifically configured to:
clear a ground truth box whose quantity of point clouds is less than the preset threshold in the plurality of ground truth boxes.
11 . The apparatus according to any one of claims 8 to 10 , wherein the millimeter-wave radar point cloud comprises a target-level point cloud and/or an original point cloud, the apparatus further comprises a motion compensation unit, and the motion compensation unit is specifically configured to:
perform motion compensation on the target-level point cloud when the millimeter-wave radar point cloud comprises the target-level point cloud.
12 . The apparatus according to any one of claims 8 to 11 , wherein the apparatus further comprises a delay compensation unit, and the delay compensation unit is specifically configured to:
perform system delay compensation on the millimeter-wave radar point cloud.
13 . The apparatus according to any one of claims 8 to 12 , wherein the apparatus further comprises a training unit, and the training unit is specifically configured to:
extract a target feature from the millimeter-wave radar point cloud, wherein the target feature comprises a plurality of the following: coordinates, static/dynamic attributes, a radar cross section, an absolute velocity, a relative velocity, a radar type, and time sequence information of a superimposed frame; and
perform training by using the plurality of ground truth boxes on which position transformation is performed and the target feature as the training dataset.
14 . The apparatus according to any one of claims 8 to 13 , wherein the millimeter-wave radar point cloud is generated by superimposing a plurality of frames of point clouds.
15 . A computer device, comprising a processor, wherein the processor is coupled to a memory; and
the processor is configured to execute instructions stored in the memory, to enable the computer device to perform the method according to any one of claims 1 to 7 .