Semantic Segmentation Data Structure for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for training semantic segmentation models using lidar or radar data face challenges due to sparsity issues and high computational complexity, leading to biased predictions that are not suitable for safety-critical applications like autonomous driving.
Innovation Solution
A computer-implemented method is introduced to create a data sample for training semantic segmentation models by combining point clouds from different times to form a global point cloud, extracting semantic and elevation maps, and using classification and bounding box annotations to separate static and dynamic objects, while reconstructing surfaces to cover uncovered areas and label vertices based on reference points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If point cloud data is used for semantic segmentation training, then the model can process 3D spatial information, but the data sparsity leads to incomplete coverage and biased predictions
Solution Approach 1:
The patent combines multiple point clouds from different time points to create a global point cloud, merging sparse data from multiple sources to achieve denser and more complete spatial coverage. This resolves the sparsity issue by aggregating information across temporal dimensions.
Solution Approach 2:
The patent transforms sparse 3D point cloud data into dense 2D occupancy grid maps by projecting 3D spatial information onto a 2D plane with multiple layers. This dimensional transformation converts sparse 3D data into dense 2D representations that preserve spatial relationships while eliminating sparsity.
2Reliability
If complex data structures with more information are used, then training performance and prediction accuracy improve, but computational cost increases significantly
Solution Approach 1:
The patent extracts only the essential spatial occupancy information from complex point cloud data and represents it in simplified grid maps. By taking out only the necessary features (occupancy status) and discarding redundant details, the method maintains training effectiveness while reducing computational complexity.
Solution Approach 2:
The patent changes the data representation parameters from raw 3D point coordinates to 2D grid cell occupancies. This parameter transformation converts complex continuous spatial data into discrete grid-based representations, reducing computational burden while preserving essential spatial relationships for semantic segmentation.
3Measurement precision
If manual annotation is used for data labeling, then data quality is high, but the effort and time required for annotation is extensive
Solution Approach 1:
The patent employs automated annotation using trained semantic segmentation models to label occupancy grid maps without manual intervention. The system self-annotates training data by applying the model to labeled data and using the predictions for further training, eliminating the need for extensive manual annotation while maintaining quality through iterative refinement.
Data Source
AI summary
Disclosed is a computer-implemented method for creating a data sample for training semantic segmentation models usable in a vehicle assistance system. The method includes obtaining a first point cloud representing a surrounding of a vehicle at a first point in time and a second point cloud representing the surrounding of the vehicle at a second point in time. The method includes joining the first and second point cloud to obtain a global point cloud representing the surrounding of the vehicle over a duration of the first point in time and the second point in time. The method includes creating a representation of the surrounding based on the global point cloud. The method includes extracting from the representation a semantic map and one or more elevation maps. The method includes providing the semantic map and the one or more elevation maps as the data sample.