LiDAR Landmark Shape Regression for Precise AV Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional perception and localization methods for autonomous vehicles using LiDAR data are limited in functionality, accuracy, and precision, particularly when detecting lane lines, which hinders effective navigation.
Innovation Solution
Utilizing deep neural networks (DNNs) to generate per-point classifications of LiDAR points and regress landmark shapes, such as lane lines and poles, through connected components analysis and shape fitting, enabling precise localization and navigation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If deep neural network inference is used to classify LiDAR points, then landmark detection capability is improved, but localization precision deteriorates due to sparse LiDAR points and scanline pattern
Solution Approach 1:
The patent segments the classification task by introducing intermediate shape representations (e.g., lane line segments, pole segments) that divide the continuous landmark structure into discrete detectable units. This segmentation allows the system to work around the sparsity of LiDAR points by fitting shapes to segmented point groups rather than requiring dense point coverage.
Solution Approach 2:
The patent introduces shape models and geometric primitives as intermediary representations between raw LiDAR points and localization outcomes. These intermediaries (e.g., fitted lane line equations, pole centerlines) bridge the gap between sparse sensor data and precise localization requirements by providing continuous geometric descriptions.
2Ease of operation
If Hough transform is used to find lines in classifications, then straight line detection is improved, but effectiveness deteriorates for curved roads and threshold selection becomes difficult
Solution Approach 1:
The patent transitions from static Hough transform assumptions (straight lines only) to dynamic shape fitting that can adapt to curved geometries. By using parametric shape models with multiple degrees of freedom (e.g., polynomial curves, splines), the system dynamically adjusts to road curvature while maintaining detection efficiency.
Solution Approach 2:
The patent changes the parameter space from simple Hough transform parameters (rho, theta) to more flexible shape description parameters (polynomial coefficients, control points). This parameter expansion allows the system to represent both straight and curved landmarks uniformly, eliminating the need for separate handling of different road geometries.
3Extent of automation
If per-point classification is used for landmark detection, then classification capability is improved, but shape regression capability deteriorates
Solution Approach 1:
The patent merges the classification output with shape regression by combining classified points into groups that form complete landmark instances. By aggregating classified points and fitting shapes to these groups, the system recovers shape information that is lost in per-point classification, achieving both automated classification and precise shape representation.
Data Source
AI summary
In various examples, perception of landmark shapes may be used for localization in autonomous systems and applications. In some embodiments, a deep neural network (DNN) is used to generate (e.g., per-point) classifications of measured 3D points (e.g., classified LiDAR points), and a representation of the shape of one or more detected landmarks is regressed from the classifications. For each of one or more classes, the classification data may be thresholded to generate a binary mask and/or dilated to generate a densified representation, and the resulting (e.g., dilated, binary) mask may be clustered into connected components that are iteratively: fitted a shape (e.g., a polynomial or Bezier spline for lane lines, a circle for top-down representations of poles or traffic lights), weighted, and merged. As such, the resulting connected components and their fitted shapes may be used to represent detected landmarks and used for localization, navigation, and/or other uses.