3D Point Cloud Model Enhancement via Image-LIDAR Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for generating three-dimensional training data using LIDAR measurements are prone to inaccuracies and omissions, leading to poor object detection in tools trained with such data.
Innovation Solution
A system that combines LIDAR distance measurements with image information to create enhanced three-dimensional point cloud models by inserting high-fidelity object models into scene models, adjusting point density and orientation based on object location and environment, to generate more accurate and detailed three-dimensional training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LIDAR measurements are used to generate three-dimensional training data, then distance information can be obtained, but measurement accuracy deteriorates due to faulty LIDAR readings
Solution Approach 1:
The patent combines LIDAR distance measurements with image information from cameras to create enhanced three-dimensional point cloud models. By merging data from multiple sources (LIDAR and image-based object models), the system compensates for faulty LIDAR readings and improves both measurement precision and reliability of object detection.
Solution Approach 2:
The patent uses image information as an intermediary to correct and enhance LIDAR data. When LIDAR readings are faulty, the system inserts high-fidelity object models derived from images into the point cloud, acting as a mediator to restore accurate distance and shape information.
2Loss of information
If physical observations of objects within scenes are used, then real-world data can be captured, but observations of particular objects/scenes are omitted
Solution Approach 1:
The patent performs preliminary action by pre-processing and storing high-fidelity object models in a database before they are needed for scene enhancement. When generating training data, the system can quickly retrieve and insert appropriate object models without needing to perform complete physical observations, thus reducing information loss while maintaining efficiency.
Solution Approach 2:
The patent creates accurate copies of objects by inserting pre-captured object models into scene point clouds. Instead of requiring complete physical observations of every object in every scene, the system uses copied object models from the database to represent objects, ensuring comprehensive object coverage without the productivity penalty of exhaustive physical observation.
3Manufacturing precision
If high-fidelity object models are inserted into scene models, then three-dimensional data accuracy improves, but processing complexity increases
Solution Approach 1:
The patent applies local quality by enhancing only specific regions of the point cloud where objects are detected, rather than processing the entire scene uniformly. The system identifies object locations using image information and applies high-fidelity models only to those local areas, improving accuracy while minimizing processing complexity.
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 fidelity of three-dimensional training data, improving object detection accuracy and enabling more effective machine learning model training for applications like vehicle control without the need for precise physical observations.
Implementation Method 1
the distances of the scene from the location may be measured using LIDAR
Data Source
AI summary
Systems, methods, and non-transitory computer readable media configured to generate enhanced three-dimensional information. Three-dimensional information of a scene may be obtained. The three-dimensional information may define a three-dimensional point cloud model of the scene. The three-dimensional information may be determined based on distances of the scene from a location. Image information may be obtained. The image information may define one or more images of an object. The object may be identified based on the image information. A three-dimensional point cloud model of the object may be obtained. Enhanced three-dimensional information of the scene may be generated by inserting the three-dimensional point cloud model of the object into the three-dimensional point cloud model of the scene.