Lidar Ground Truth Alignment for Millimeter-Wave Radar Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in the low matching degree between laser ground truth and millimeter-wave radar point clouds due to the different working principles of lidar and millimeter-wave radar, leading to inaccurate training of target detection models.
Innovation Solution
A data processing method is provided that involves obtaining lidar and millimeter-wave radar point clouds, performing position transformation on ground truth boxes until a preset condition is met, and using the transformed ground truth boxes and millimeter-wave radar point clouds as a training dataset to optimize the target detection model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If laser ground truth is used to annotate millimeter-wave radar point clouds, then annotation efficiency is improved, but detection accuracy deteriorates due to position mismatch between lidar and millimeter-wave radar point clouds
Solution Approach 1:
The patent uses lidar point cloud annotations as a reference copy, then transforms and adapts them to match millimeter-wave radar point cloud positions. The ground truth boxes are copied from lidar data and undergo position transformation to align with the actual millimeter-wave radar detection positions, resolving the position mismatch while maintaining annotation efficiency.
Solution Approach 2:
The patent transforms the position parameters of ground truth boxes by calculating offset values between lidar and millimeter-wave radar point cloud positions. These parameter changes (position adjustments) are applied to the ground truth boxes to make them consistent with millimeter-wave radar detection data, thereby improving detection accuracy while preserving the efficiency of using lidar annotations.
2Measurement precision
If position transformation is performed on ground truth boxes to improve matching degree, then detection accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent performs position transformation on ground truth boxes in advance during the data preparation phase, before the actual detection model training. By pre-calculating and adjusting the positions of ground truth boxes to match millimeter-wave radar point clouds, the complexity is handled upfront, allowing the detection model training to proceed with already-optimized data.
Solution Approach 2:
The patent introduces an intermediate processing step that acts as a mediator between lidar annotations and millimeter-wave radar data. The position transformation process serves as an intermediary mechanism that translates ground truth boxes from lidar coordinate system to millimeter-wave radar coordinate system, simplifying the overall integration process while improving accuracy.
Data Source
AI summary
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.