LIDAR Point Cloud Filtering for More Accurate Dataset Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for merging 3D point clouds from LIDAR sensors in autonomous vehicles face challenges in achieving accurate convergence and object detection due to limited error metric minimization, leading to distorted representations and reduced accuracy in object detection.
Innovation Solution
A modified Iterative Closest Point (ICP) framework that includes initialization, filtration, matching, transformation estimation, and termination criteria steps, utilizing Normal Covariance Filter (NCF) for data point filtration and geometric criteria for outlier rejection, to improve the accuracy of merging 3D point clouds and enhance object detection in autonomous vehicles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the standard ICP algorithm is used to merge 3D point clouds, then the merging process can be completed, but the convergence accuracy is limited and error metric minimization is insufficient leading to distorted representations
Solution Approach 1:
The patent applies preliminary action by performing data filtration before the ICP merging process. Specifically, it filters out erroneous LIDAR data points using statistical methods (such as removing points beyond a certain standard deviation threshold from the mean depth) and geometric consistency checks before merging. This preliminary cleaning of data ensures that the ICP algorithm works with higher quality input, thereby improving convergence accuracy and reducing distortion in the merged point cloud representation.
Solution Approach 2:
The patent changes parameters by modifying the error metric minimization approach in the ICP algorithm. It introduces weighted error metrics that prioritize certain regions or types of points over others, and adjusts convergence thresholds dynamically. These parameter changes allow the algorithm to achieve better convergence accuracy and produce more reliable merged representations suitable for accurate object detection.
2Area of stationary object
If multiple LIDAR datasets are merged to create a comprehensive scene representation, then coverage is improved, but computational load increases and processing time is extended
Solution Approach 1:
The patent applies the taking out principle by extracting and removing erroneous or redundant data points from the LIDAR datasets before merging. It uses statistical filtering to identify and eliminate outliers, and geometric consistency checks to remove inconsistent points. This extraction of bad data reduces the overall dataset size and complexity, thereby decreasing computational load and processing time while maintaining comprehensive scene coverage from multiple LIDAR sources.
3Manufacturing precision
If data filtration is applied to remove erroneous points, then merging accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the data filtration process into multiple distinct stages: first performing statistical filtering based on depth distribution, then applying geometric consistency checks, and finally using outlier detection algorithms. Each stage handles a specific aspect of error removal with relatively simple operations. This segmented approach improves merging accuracy through comprehensive filtration while keeping individual processing steps computationally manageable, avoiding the need for a single complex filtration algorithm.
Data Source
AI summary
Method and device for processing LIDAR sensor data are disclosed. The method includes (i) receiving from the LIDAR sensor a first dataset having a plurality of first data points representative of respective coordinates and associated with respective normal vectors, (ii) determining an uncertainty parameter for a given first data point based on a normal covariance of the normal vector of the given first data point where the normal covariance takes into account a measurement error of the LIDAR sensor when determining the respective coordinates of the given first data point, (iii) in response to the uncertainty parameter being above a pre-determined threshold, excluding the given first data point from the plurality of first data points, (iv) using the filtered plurality of first data points, instead of the plurality of first data points, for merging the first dataset of the LIDAR sensor with a second dataset of the LIDAR sensor.