Lane Segment Detection in BEV for Intersection HD Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for creating high-definition (HD) maps struggle to accurately predict lane segments, especially in intersections, leading to unwanted gaps and failures in detecting curved segments, which hampers the safe navigation of autonomous vehicles.
Innovation Solution
A two-step neural network approach is employed, where a first neural network predicts pixel properties for boundary lines, lanes, and driving directions, followed by a second network treating roads and lanes as a graph to connect lane segments, especially in intersections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network models operate in image space to detect lane segments, then they can predict straight lane segments, but they fail to detect lane segments in intersections with overlapping lanes
Solution Approach 1:
The patent transforms the lane detection problem from 2D image space to 3D bird's eye view (BEV) space by applying geometric projection and transformation. This dimensional change allows the neural network to perceive lane segments in intersections more accurately, as the BEV representation naturally accommodates overlapping and curved lane configurations that are difficult to detect in traditional 2D image space.
Solution Approach 2:
The patent changes the operational parameters of the neural network by inputting transformed features from multiple camera views and applying geometric adjustments. The network processes images through a series of transformations including perspective projection, rotation, and scaling to convert 2D camera images into 3D BEV representations, enabling accurate detection of curved and overlapping lane segments in intersections.
2Ease of manufacture
If classical object detection and segmentation models are used, then processing is straightforward, but gaps appear between adjacent lane segments in intersections
Solution Approach 1:
The patent segments the lane detection task into multiple stages: first detecting lane segments in individual camera images, then projecting them to BEV space, and finally merging adjacent segments based on spatial proximity and geometric consistency. This segmentation approach allows complex intersection scenarios to be handled systematically, ensuring continuous and accurate lane segment representation.
Solution Approach 2:
The patent introduces BEV space as an intermediary representation between 2D camera images and the final lane segment output. This intermediate 3D representation serves as a mediator that reconciles the simplicity of 2D processing with the precision requirements of 3D lane segmentation, enabling accurate merging of adjacent segments while maintaining computational efficiency.
3Productivity
If automated HD map creation processes are used, then time and costs are reduced, but accuracy in detecting curved and overlapping lane segments deteriorates
Solution Approach 1:
The patent implements a self-service automated system where the neural network processes multiple camera images, performs geometric transformations, detects lane segments, and merges them into HD maps autonomously. The system uses its own extracted features and transformations to generate the final map without requiring external intervention, achieving both high productivity and maintained precision through iterative self-processing.
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for detecting lane segments in an image for creating high-definition (HD) maps. A neural network can be used to classify a pixel of an image of a location as belonging to a driving lane. If the pixel belongs to a first boundary line of the driving lane, it is labeled accordingly. Based on the labeled pixel and one or more additional labeled pixels as part of the first boundary line of the driving lane, a first line drawing of the first boundary line of the driving lane is constructed. A lane segment based on a combination of the first line drawing of the first boundary line of the driving lane and a second line drawing of a second boundary line of the driving lane can be created.