Point Cloud Noise Reduction via Kernel Filtering and Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating point cloud data from real target objects introduce noise, particularly in depth calculations, which affects the accuracy of segmentation when used as teacher data for network training. Additionally, generating teacher data for flexible elastic objects is challenging using CAD data, and existing noise removal approaches can incorrectly identify points on distant objects as noise.
Innovation Solution
A method that involves capturing two-dimensional color image data of a target object, classifying pixels into types, determining kernels based on region sizes, generating point cloud data, projecting points onto a plane, associating points with classified pixels, and filtering point cloud data using kernels to reduce noise and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If point cloud data is generated from real target objects using LiDAR and stereo camera, then the data reflects actual target objects, but noise is introduced particularly in depth direction calculations
Solution Approach 1:
The patent introduces a teacher network as an intermediary that processes the noisy point cloud data. The teacher network learns to distinguish between actual target object points and noise points by training on labeled data, thereby mediating between the noisy sensor data and the clean segmentation results needed for accurate depth calculation and object recognition
Solution Approach 2:
The patent segments the point cloud data into different categories: points belonging to target objects and noise points. By classifying and separating these different types of points through the trained network, the system can process only the relevant target object points for segmentation tasks, effectively removing noise without losing important geometric information
2Ease of manufacture
If a fixed density threshold is used to remove low-density point clouds, then noise removal is simplified, but points on distant target objects are incorrectly removed as noise
Solution Approach 1:
The patent applies different processing characteristics to different regions of the point cloud data. Instead of using a uniform density threshold, the system evaluates each point's local neighborhood characteristics and classification probability independently, allowing distant objects with naturally lower density to be preserved while still removing actual noise points based on their local structural properties
Solution Approach 2:
The patent changes from using a fixed density parameter for noise removal to using a learned classification probability parameter. The teacher network outputs probability values that dynamically adjust the effective density threshold for each point based on its contextual features, enabling adaptive noise removal that preserves distant objects while removing noise
3Measurement precision
If teacher data is generated from CAD data for each target object, then accurate segmentation training is achieved, but the burden of data generation becomes large
Solution Approach 1:
The patent uses a teacher network that has been trained on a comprehensive dataset to create a generalized model that can process new target objects without requiring manual CAD-based teacher data generation for each object. The trained network copies the segmentation knowledge learned from training data and applies it automatically to new objects, eliminating the time-consuming CAD data processing step
Solution Approach 2:
The patent performs preliminary training of the teacher network on a large dataset of labeled point clouds and CAD models in advance. This preliminary action creates a pre-trained model that encapsulates segmentation knowledge, which can then be applied to new objects without requiring repeated CAD data processing and manual labeling for each new target object
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
The proposed method effectively reduces noise in point cloud data, improves the accuracy of segmentation, and allows for the generation of accurate point cloud data for flexible elastic objects without relying on CAD data, thereby enhancing the performance of point cloud processing systems.
Implementation Method 1
a time-of-flight (ToF) camera, such as LiDAR, radially emits light and generates point cloud data based on the period spent until the camera receives the reflected light
Data Source
AI summary
Color image data is generated by capturing an image of a target object. Pixels of the color image data are classified to generate classified image data. The kernels have sizes according to a plurality of sub-target objects formed of pixels that belong to categories. First data on a point cloud is generated by capturing an image of the target object using an infrared stereo camera. The first data is projected onto a plane to generate second data on a point cloud. A point of the point cloud in the second data is associated with a pixel in the classified image data. Third data on a point cloud in a three-dimensional coordinate system is generated from the first data by filtering information on the positions of the points in the first data. The filtering uses the kernels according to the categories of the points contained in the first data.