LiDAR Map Annotation via Key Frame Selection and Trajectory Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing annotation methods for LiDAR data are inefficient and lack consistency, leading to repetitive processes and affecting training efficiency, with annotation results potentially being reset on subsequent frames.
Innovation Solution
A method and computing device that selectively choose key frames for annotation, allowing simultaneous display of LiDAR maps and camera images, and cumulatively record annotation data to generate an annotated LiDAR map, using a LiDAR SLAM algorithm to generate key frame trajectories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If annotation is performed per frame for all LiDAR frames, then complete annotation coverage is achieved, but annotation time and computational resources are excessively consumed due to repetitive processes
Solution Approach 1:
The patent segments the continuous stream of LiDAR frames into discrete key frames based on significant changes in scene content or vehicle pose. Instead of annotating every frame, the system identifies and selects only those frames that contain meaningful changes, thereby reducing the total number of annotation tasks while maintaining comprehensive coverage of all annotated objects across the trajectory.
Solution Approach 2:
The system performs preliminary processing to generate the LiDAR map and pre-identify key frames before the actual annotation process. By preparing the annotated LiDAR map in advance and selecting key frames based on predetermined criteria (such as pose change thresholds or detection of new objects), the system eliminates the need for repetitive annotation on every frame while ensuring all necessary objects are captured.
2Manufacturing precision
If annotation is performed on every frame independently, then each frame is fully annotated, but consistency of annotation across frames deteriorates due to reset issues in annotation tools
Solution Approach 1:
The patent establishes a continuous annotation process where annotations are accumulated across key frames rather than being reset for each frame. The annotated LiDAR map serves as a persistent data structure that maintains annotation state throughout the trajectory, allowing annotators to continue labeling objects across multiple key frames without losing previous annotation context or having to re-annotate already-labeled objects.
Solution Approach 2:
The system provides feedback mechanisms where the annotated LiDAR map is continuously updated and reflected across all key frames. When an object is annotated in one key frame, the annotation information is propagated and made available in subsequent key frames, allowing annotators to verify consistency and make adjustments if needed. This feedback loop ensures that annotation decisions are maintained consistently across the entire trajectory.
3Quantity of substance
If all LiDAR frames are annotated, then comprehensive training data is generated, but the complexity of managing and processing annotation data increases significantly
Solution Approach 1:
The patent extracts only the essential annotation information from the LiDAR frames by selecting key frames that contain significant changes. Instead of processing and storing annotation data for every frame, the system extracts and stores annotations only for key frames, significantly reducing the volume of data that needs to be managed while still providing sufficient training data coverage for machine learning models.
Solution Approach 2:
The system transitions from frame-by-frame annotation in temporal dimension to trajectory-based annotation in spatial dimension. By organizing annotations around the vehicle's trajectory and key positional changes rather than chronological frame sequence, the patent reduces data management complexity while maintaining comprehensive coverage of all annotated objects throughout the driving route.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
A method for annotation based on a LiDAR map is provided. The method includes steps of: (a) on condition that each of datasets has been recorded in a database, in response to acquiring a specific dataset, generating a specific LiDAR map and a - specific key frame trajectory using a plurality of specific LiDAR point cloud data included in the specific dataset, wherein each of the datasets includes (1) each of LiDAR point cloud data corresponding to each of LiDAR frames and (2) each of camera images corresponding to each of camera frames, acquired by a predetermined criterion while driving each of the driving routes; and (b) allowing a specific key frame, which is at least part of a plurality of key frames included in the specific key frame trajectory, to be annotated, thereby generating an annotation result and thus recording the annotation result in the specific LiDAR map.