Graph-Based BEV Perception for Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing camera-based driving environment perception technologies using segmentation are limited by increased data heaviness as the perception area and map size grow, making them unsuitable for comparison with high-definition (HD) maps and hindering the advancement of autonomous driving, especially for level 4/4+ autonomous driving and global scalability.
Innovation Solution
A graph-based driving environment perception method that detects and models driving environment objects in a vector form, transforming them into a graph form using nodes and edges to represent connectivity, enabling efficient perception and localization compatible with HD maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If segmentation technology is used for driving environment perception, then perception capability is improved, but data weight increases making it unsuitable for HD map comparison
Solution Approach 1:
The patent segments the driving environment into distinct semantic categories (road, sidewalk, building, vegetation, sky, other) and represents each category with lightweight vector primitives (points, lines, polygons) rather than heavy raster segmentation masks. This segmentation approach maintains perception capability while significantly reducing data weight for HD map comparison.
Solution Approach 2:
The patent transforms the representation parameters from dense raster grid values to sparse vector geometric parameters (coordinates, connectivity relationships). This parameter change reduces data weight while preserving essential spatial and semantic information needed for perception and HD map matching.
2Area of stationary object
If rasterization of BEV space is used for segmentation, then perception coverage is improved, but computational complexity increases
Solution Approach 1:
Instead of rasterizing the entire BEV space and then segmenting, the patent inverts the approach by directly detecting and representing environmental elements as vector primitives in BEV space. This avoids the computational burden of full-space rasterization while maintaining comprehensive perception coverage.
Solution Approach 2:
The patent extracts only the essential geometric and semantic information needed for perception (vertices, edges, polygons, connectivity) from the complex BEV scene, rather than processing the entire rasterized space. This extraction approach reduces computational complexity while preserving perception coverage.
3Quantity of substance
If vector form representation is used for driving environment objects, then data lightweight is improved, but connectivity information is lost
Solution Approach 1:
The patent nests connectivity information within the vector representation structure by organizing geometric primitives (points, lines, polygons) into hierarchical relationships where connectivity is encoded through parent-child relationships and adjacency matrices. This nesting preserves connectivity information while maintaining data lightweight characteristics.
Solution Approach 2:
The patent adds a connectivity dimension to the traditional 2D geometric representation by introducing graph-based relationships between vector primitives. This dimensional enhancement preserves topological and connectivity information without significantly increasing data weight, enabling both lightweight representation and connectivity awareness.
Data Source
AI summary
Disclosed is a method and system for graph-based bird's-eye-view (BEV) driving environment perception for autonomous driving. A graph-based driving environment perception method for autonomous driving may include detecting driving environment objects on the road in a vector form; and modeling the driving environment objects from the vector form to a graph representation.