LiDAR Object Labeling via Trajectory Refinement for Occlusion Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual annotation of datasets for self-driving systems is a bottleneck, especially in handling object occlusions, sparsity of observations, and diverse size and motion profiles, which hampers the training of object detectors.
Innovation Solution
Implementing automatic labeling of objects from LiDAR point clouds via trajectory level refinement using an encoder model, attention model, and decoder model to generate refined bounding box vectors, which address occlusions, sparsity, and diversity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to label objects in LiDAR point clouds, then labeling accuracy can be maintained, but productivity is severely limited and costs increase
Solution Approach 1:
The system uses automatically generated bounding box vectors from trajectory models as initial labels, which then serve as input for refinement models. This self-service approach allows the system to generate its own training labels without human intervention, dramatically improving productivity while maintaining accuracy through iterative refinement processes
Solution Approach 2:
The patent replaces the mechanical process of manual annotation with automated machine learning models (encoder models, attention models, and decoder models). These computational systems process LiDAR point clouds and generate refined bounding boxes automatically, eliminating the bottleneck of human annotators while maintaining or improving labeling quality
2Productivity
If automatic labeling is implemented without trajectory refinement, then productivity improves, but handling of object occlusions and sparsity of observations deteriorates
Solution Approach 1:
The trajectory model generates preliminary bounding box vectors before the refinement stage. These preliminary labels incorporate temporal information and motion predictions that pre-compute potential object positions even when occluded, providing a head start for handling occlusions and sparse observations before the refinement process begins
Solution Approach 2:
The refinement model uses feedback from multiple sources including original LiDAR observations, trajectory predictions, and previously generated bounding boxes to iteratively improve label accuracy. This feedback loop allows the system to correct errors in preliminary annotations and handle difficult cases like occlusions and sparse observations more effectively
3Speed
If simple automatic labeling is used, then processing speed increases, but accuracy in diverse size and motion profiles deteriorates
Solution Approach 1:
The labeling system is segmented into distinct functional components: trajectory models for motion prediction, encoder models for feature extraction, attention models for selective focus, and decoder models for final label generation. This segmentation allows each component to be optimized for its specific task while working together to achieve both speed and accuracy for diverse object types
Solution Approach 2:
The system employs dynamic models that adapt to diverse object characteristics. The trajectory models learn different motion profiles for various object types (vehicles, pedestrians, cyclists), and the refinement process dynamically adjusts based on object size, speed, and behavior patterns, maintaining high accuracy across diverse size and motion profiles while preserving processing efficiency
Data Source
AI summary
A method implements automatic labeling of objects from LiDAR point clouds via trajectory level refinement. The method includes executing an encoder model using a set of bounding box vectors and a set of point clouds to generate a set of combined feature vectors and executing an attention model using the set of combined feature vectors to generate a set of updated feature vectors. The method further includes executing a decoder model using the set of updated feature vectors to generate a set of pose residuals and a size residual and updating the set of bounding box vectors with the set of pose residuals and the size residual to generate a set of refined bounding box vectors. The method further includes executing an action responsive to the set of refined bounding box vectors.