Autonomous Vehicle Localization Using Walsh Kernel Map Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle localization methods, particularly those using LIDAR sensors, are computationally expensive and lack robustness and accuracy, especially in dynamic urban environments with complex signal occlusions, failing to achieve the required precision of within 10 cm for high-definition map positioning.
Innovation Solution
A computer-implemented method for autonomous driving vehicles that determines location by breaking down 3D point cloud data into cells, generating a feature space with mean intensity and variance, and using a similarity metric based on Gray Code Kernels to align the vehicle's feature space with a high-definition map, allowing for efficient and accurate positioning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 3D point cloud data from LIDAR sensor is used for autonomous vehicle localization, then positioning accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent divides the 3D point cloud data into multiple two-dimensional projections (top-down view and side views). This segmentation transforms the complex 3D localization problem into simpler 2D matching problems, reducing computational complexity while maintaining positioning accuracy through multi-view geometry integration.
Solution Approach 2:
The patent introduces Gray Code Kernels as an intermediary computational tool to bridge the LIDAR point cloud data and the HD map features. These kernels enable efficient similarity computation between sensor data and map data through mathematical transformation, reducing the direct computational burden of 3D point cloud processing.
2Measurement precision
If traditional LIDAR-based localization methods are used, then positioning accuracy is improved, but processing speed decreases
Solution Approach 1:
The patent employs periodic scanning patterns of the LIDAR sensor to systematically collect point cloud data. This periodic action, combined with real-time projection and matching algorithms, enables continuous localization updates at high frequency, improving processing speed while maintaining accuracy through regular data acquisition cycles.
Solution Approach 2:
By segmenting the 3D point cloud into multiple 2D projections that can be processed independently and in parallel, the system achieves faster processing speeds. The computational tasks are divided into manageable units (different view projections) that can be handled simultaneously, reducing overall processing time while preserving positioning accuracy.
3Reliability
If existing localization methods are used in dynamic urban environments with signal occlusion, then robustness should be improved, but computational cost increases
Solution Approach 1:
The patent transitions from 3D point cloud processing to 2D projection processing, effectively changing the dimensionality of the problem. This dimensional reduction simplifies the computational complexity while maintaining robustness in occluded environments, as the top-down and side view projections provide complementary information that compensates for signal blockages from any single viewpoint.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Location of an autonomous driving vehicle (ADV) is determined with respect to a high definition map. On-boards sensors of the ADV obtain a 3D point cloud of objects surrounding the ADV. The 3D point cloud is organized into an ADV feature space of cells. Each cell has a median intensity value and a variance in elevation. A set of candidate cells that surround the ADV is determined. For each candidate, a set of cells of the ADV feature space that surround the candidate cell is projected onto the map feature space using kernel projection, for one or more dimensions. Kernels can be Walsh-Hadamard vectors. Candidates having insufficient similarity are rejected. When a threshold number of non-rejected candidates remain, candidate similarity can be determined using a similarity metric. The coordinates of the most similar candidate cell are used to determine the position of the vehicle with respect to the map.