LiDAR Ground Segmentation via Local Adaptive Region Proposals
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for ground surface segmentation in LiDAR-based vehicle navigation systems are inefficient due to reliance on simplistic plane assumptions, high computational costs, and dependency on additional sensors or deep learning models, making them unsuitable for real-time applications.
Innovation Solution
A ground point segmentation algorithm that computes a piece-wise local ground representation using a maximum height map and minimum filtering, eliminating the need for prior knowledge of LiDAR type or additional sensors, and reducing computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional plane fitting methods (RANSAC) are used for ground segmentation, then ground representation can be obtained, but computational cost becomes significantly high
Solution Approach 1:
The patent divides the ground segmentation problem into two stages: (1) coarse segmentation using a simplified method to identify candidate ground regions, and (2) refined segmentation applying more accurate methods only to these reduced regions. This hierarchical segmentation approach maintains accuracy while dramatically reducing computational cost by avoiding exhaustive processing of the entire point cloud.
Solution Approach 2:
The patent applies the principle of partial action by using a lightweight ground estimation method to process the entire point cloud first, then applying computationally intensive refinement only to specific regions of interest. This selective application of processing intensity optimizes the balance between speed and accuracy.
2Productivity
If a single plane assumption is used for ground representation, then computational complexity is reduced, but performance deteriorates in sophisticated real-world scenes
Solution Approach 1:
The patent segments the ground surface into multiple local planar regions rather than assuming a single global plane. By dividing the scene into manageable patches and fitting planes locally, the method captures complex ground geometries while keeping individual computations simple and efficient.
Solution Approach 2:
The patent applies different levels of ground representation to different spatial regions. In areas with simple ground surfaces, a single plane suffices, while in complex regions, multiple local planes are used. This local adaptation of representation quality optimizes both accuracy and computational efficiency.
3Measurement precision
If deep learning models are used for ground segmentation, then segmentation accuracy improves, but computational cost and time increase making them unsuitable for real-time applications
Solution Approach 1:
The patent introduces geometric reasoning and height-map-based intermediate representations as mediators between raw LiDAR data and final ground segmentation. These intermediaries provide structurally informed estimates that guide subsequent processing, reducing reliance on computationally expensive deep learning while maintaining accuracy through physics-based constraints.
Solution Approach 2:
The patent replaces data-driven deep learning approaches with geometry-driven algorithms based on physical principles of light reflection and surface geometry. This substitution uses deterministic geometric reasoning instead of probabilistic neural network inference, achieving real-time performance through analytically solvable equations.
4Measurement precision
If complex geometric reasoning or stereo camera approaches are used, then ground segmentation accuracy improves, but device complexity and computational requirements increase
Solution Approach 1:
The patent develops a ground segmentation method that works universally across different LiDAR types (rotating and solid-state) without requiring type-specific calibration or additional sensors. The approach uses only the intensity and spatial information inherently available from standard LiDAR systems, making it broadly applicable while avoiding increased device complexity.
Solution Approach 2:
The patent enables the LiDAR system to segment ground using only its own captured data (point cloud coordinates and intensity), without requiring external sensors like stereo cameras or prior knowledge about LiDAR-scene interaction. The method extracts all necessary information from the LiDAR return itself, making the system self-sufficient.
Data Source
AI summary
Embodiments include a method for object detection in a Light Detection And Ranging (LiDAR) point cloud, the method comprising: placing, by a navigation system, a plurality of anchor points in a two-dimensional Bird's Eye View (BEV) of spatial points represented in a segmented ground surface representation of objects detected by a LiDAR system; extracting, by the navigation system, one or more features from the two-dimensional BEV of the spatial points; proposing, by the navigation system, one or more regions of the two-dimensional BEV of the spatial points for object detection; and performing, by the navigation system, object detections on anchor points of the plurality of anchor points in the proposed one or more regions of the two-dimensional BEV of the spatial points.