Compressed Graph HD Maps for Cross-Region Generalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing HD map generation techniques result in large map sizes and lack generalizability across regions, requiring significant computational resources and failing to adapt to out-of-distribution data, leading to outdated dynamic features and complex navigation strategies.
Innovation Solution
A reconfigurable graph-map network with long-term graph embeddings is used to generate compressed HD maps by refining graph edges based on active learning and out-of-distribution trace-activations, incorporating text-graph correspondences and token quantization to preserve joint attention and enhance generalizability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If HD maps are generated using traditional sensor fusion methods, then map precision is improved, but map size becomes very large (reaching couple of terabytes per day)
Solution Approach 1:
The patent extracts only the essential semantic information and structural features from the complete sensor data, rather than storing all raw data. This is achieved by identifying and retaining key map elements (lanes, intersections, traffic signs) while discarding redundant information, thereby reducing map size while preserving necessary precision for navigation
Solution Approach 2:
Instead of generating complete high-precision maps and then compressing them, the patent inverts the approach by directly generating compressed representations that capture only the essential semantic information needed for autonomous navigation, achieving both compression and usability from the outset
2Measurement precision
If machine-learning models are trained on specific region data, then model accuracy for that region is improved, but generalizability to other regions deteriorates
Solution Approach 1:
The patent creates a universal map representation framework that can handle multiple regions and scenarios with a single system. The semantic map structure is designed to be region-agnostic, using standardized elements (lanes, intersections, traffic controls) that apply across different geographical areas, enabling the same model to generalize to out-of-distribution samples
3Quantity of substance
If dynamic features are stored centrally on cloud servers, then data storage capacity is improved, but navigation response time deteriorates
Solution Approach 1:
The patent segments the map data into static components (road geometry, infrastructure) and dynamic components (traffic conditions, obstacles). Static components can be pre-processed and stored efficiently, while dynamic components are updated selectively and transmitted only when changes occur, reducing both storage requirements and communication latency for navigation decisions
4Loss of information
If multiple map layers are used to capture comprehensive environment information, then information completeness is improved, but computational complexity deteriorates
Solution Approach 1:
The patent merges multiple semantic map layers (road geometry, traffic rules, dynamic obstacles, environmental context) into a unified hierarchical representation. This integrated structure allows the system to access comprehensive information through a single coherent model rather than managing separate layers, reducing computational overhead while maintaining information completeness
Data Source
AI summary
In one embodiment, a method includes accessing sensor data captured by mobile devices operating in multiple regions in an environment, generating graph embeddings associated with each mobile device for each region based on a compressed graph constructed from the sensor data, generating a refined graph for each region based on the graph embeddings associated with each mobile device by reconfiguring edges in the compressed graph, generating a graph high-definition (HD) map associated with each mobile device based on a fusion of the refined graphs, identifying prominent nodes and edges connecting the prominent nodes based on the graph HD map associated with each mobile device based on trace activations associated with the prominent edges, and generating a compressed graph HD map for the environment based on the prominent nodes and edges associated with the mobile devices and environmental information associated with the environment.