Car Radar Data Annotation via Co-Registered Camera Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automobile radar datasets lack comprehensive annotations, particularly for objects hidden behind others, and existing annotation methods are costly and prone to errors, limiting the performance of machine learning algorithms.
Innovation Solution
A system and method that utilize a combination of radar and optical detection systems, such as cameras or stereo cameras, to generate and semantically segment camera images, align them with radar images in a common coordinate system, and automatically annotate radar targets based on the segmented object classes in the camera image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation methods are used to annotate radar targets, then annotation accuracy can be maintained, but time consumption and costs increase significantly
Solution Approach 1:
The system enables automated annotation where the annotation process serves itself by using the semantic segmentation results to directly label radar targets without requiring manual intervention for each target, thus achieving both high accuracy and efficiency
Solution Approach 2:
Semantic segmentation of camera images acts as an intermediary that bridges radar detection and annotation, transferring object class information from the optical domain to the radar domain through coordinate system transformation and matching
2Quantity of substance
If radar point clouds are used for annotation, then storage requirements are reduced, but information completeness is lost
Solution Approach 1:
The system transitions from analyzing radar point clouds in three-dimensional space to using two-dimensional semantic segmentation maps from camera images, leveraging the different dimensional strengths of each modality to achieve complete annotation
3Loss of information
If camera images are used for annotation, then object class information is obtained, but distance information and hidden objects are lost
Solution Approach 1:
The system merges the strengths of both radar and camera systems by combining semantic segmentation from camera images with target detection and distance measurement from radar, creating a comprehensive annotation that includes both object class and precise spatial information
4Adaptability or versatility
If existing public radar datasets are used, then data availability is improved, but annotation completeness and quality are insufficient
Solution Approach 1:
The system performs preliminary semantic segmentation of camera images and pre-aligns them with radar data in a common coordinate system, preparing complete and accurate annotations in advance that can be directly applied to radar targets, thereby ensuring high annotation quality from the outset
Data Source
AI summary
The invention relates to a system for annotating car radar data, comprising: at least one radar arranged on a car for producing a radar image by means of radar measurement; at least one optical detection system arranged outside the car for producing a camera image; a segmentation unit, which is designed to subject a camera image produced by the optical detection system to semantic segmentation for forming a semantic grid in order to assign one of a plurality of object classes to the camera image pixel by pixel; a computing unit, which is designed to transfer the camera image and/or the radar image into a common coordinate system for co-registration; and an annotation unit, which is designed to carry out annotation of the radar image, in other words to allocate an object class to a radar target of the radar image, in such a way that the object class of the semantic grid of the co-registered camera image in which the radar target of the co-registered radar image is located is allocated to a particular radar target.