Point Cloud Registration via Coarse Alignment for AV Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for point cloud registration in autonomous vehicles, such as the Iterative Closest Point (ICP) algorithm, are slow and error-prone due to inaccurate initial GPS estimates, leading to inaccuracies in determining the vehicle's position and orientation relative to HD maps.
Innovation Solution
A method involving deriving a query point cloud from lidar sensor data and a reference point cloud from an HD map, extracting features, determining a coarse alignment, and refining it using an ICP algorithm to estimate the vehicle's position-orientation pose, which improves the accuracy and speed of point cloud registration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ICP algorithm is used for point cloud registration, then positioning accuracy can be improved, but the process becomes slow and error-prone due to inaccurate initial GPS estimates
Solution Approach 1:
The patent applies preliminary action by performing feature extraction and coarse alignment before the ICP algorithm execution. Specifically, it extracts features from both query and reference point clouds, performs initial matching to obtain a rough transformation matrix, and uses this pre-aligned result as input for ICP. This preliminary preparation eliminates the need for ICP to start from an inaccurate GPS estimate, thereby improving both accuracy and speed.
Solution Approach 2:
The patent segments the point cloud registration process into distinct stages: feature extraction, coarse alignment through feature matching, and refinement via ICP algorithm. By dividing the overall registration task into these sequential steps, each stage can be optimized independently, with the feature matching providing a good initial estimate that enables the ICP to converge faster and more reliably.
2Reliability
If multiple ICP algorithms are run with different seed poses to improve accuracy, then positioning reliability increases, but the computational complexity and time consumption increase significantly
Solution Approach 1:
The patent performs preliminary feature extraction and coarse alignment before ICP execution, creating a reliable initial estimate that enables single-run ICP convergence. This preliminary preparation eliminates the need for multiple ICP trials with different seed poses, achieving high reliability through a single optimized registration process rather than multiple attempts.
Solution Approach 2:
The patent replaces the mechanical trial-and-error approach of running multiple ICP algorithms with different seed poses by substituting it with a feature-based coarse alignment system. Instead of mechanically repeating ICP with varied inputs, the system uses feature extraction and matching to compute an initial transformation matrix that guides the ICP algorithm to converge correctly in a single execution.
3Ease of operation
If GPS measurements are used as initial alignment estimate, then the bootstrap process can start, but the inaccurate GPS data (off by multiple meters) causes registration inaccuracies
Solution Approach 1:
The patent introduces feature matching as an intermediary step between GPS-based coarse localization and ICP-based precise alignment. The feature extraction and matching process acts as a mediator that refines the inaccurate GPS estimate into an accurate initial alignment estimate, enabling the ICP algorithm to converge to the correct pose without being misled by GPS errors of multiple meters.
Solution Approach 2:
The patent performs preliminary feature-based coarse alignment to correct GPS inaccuracies before initiating the ICP process. By extracting features from both point clouds and computing initial matches, the system prepares an improved initial estimate that eliminates the multiple-meter GPS errors, allowing the bootstrap process to proceed with high accuracy from the outset.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach enhances the accuracy and speed of point cloud registration, reducing errors and convergence time, enabling more reliable autonomous vehicle navigation.
Implementation Method 1
deriving, by one or more computing devices, a query point cloud from a sweep of a light detection and ranging (lidar) sensor device of a vehicle
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for fast point cloud registration for bootstrapping localization in a high-definition (HD) map. A computing device in an autonomous vehicle (AV) derives a query point cloud from a sweep of a light detection and ranging (lidar) sensor device of the AV and a reference point cloud from an HD map. The computing device extracts a first set of features from the query point cloud and a second set of features from the reference point cloud. The computing device calculates a coarse alignment based on a plurality of matches between the first set of features from the query point cloud and the second set of features from the reference point cloud. Finally, the computing device estimates the position-orientation pose of the vehicle by refining the coarse alignment using an iterative closest point (ICP) algorithm.