Real-Time Landmark Extraction from Sparse 3D Point Clouds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting landmarks from 3D point clouds are computationally intensive and unsuitable for autonomous vehicles due to their reliance on dense point clouds and offline processing, which are not feasible with the limited resources available on these vehicles, especially when dealing with sparse point clouds generated by onboard LIDAR sensors.
Innovation Solution
The method involves converting sparse 3D point clouds into range maps, labeling data points as pole-like or vertical-plane-like, and generating segmented point clouds without labeling every data point, leveraging the characteristics of LIDAR sensors such as beam numbers and Azimuth angles to reduce computational complexity and distortions, effectively classifying ground and ceiling points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing methods for extracting landmarks from 3D point clouds are used, then landmark extraction accuracy is improved, but computational complexity increases making it unsuitable for autonomous vehicles
Solution Approach 1:
The patent segments the point cloud processing by dividing it into two distinct stages: an offline training phase using dense point clouds to build classification models, and an online extraction phase using sparse point clouds with the pre-trained models. This segmentation allows complex computations to be performed once offline, while real-time processing uses lightweight models, thereby resolving the contradiction between extraction accuracy and computational complexity for autonomous vehicles.
Solution Approach 2:
The patent applies preliminary action by performing offline training and model construction before real-time landmark extraction. Dense point clouds are processed in advance to train classification algorithms and build feature extractors. These pre-trained models are then deployed for efficient real-time processing of sparse point clouds from autonomous vehicle LIDAR sensors, eliminating the need for complex real-time computations while maintaining high accuracy.
2Measurement precision
If dense 3D point clouds are used for landmark extraction, then extraction accuracy is improved, but resource requirements increase making it infeasible for autonomous vehicles
Solution Approach 1:
The patent uses copying by creating synthetic dense point cloud data from sparse real-world LIDAR measurements through data augmentation techniques. Virtual dense point clouds are generated by interpolating and expanding sparse measurements, then used to train classification models. This allows the system to learn from dense data patterns without requiring actual dense point clouds during real-time autonomous vehicle operation, thus maintaining accuracy while reducing resource requirements.
3Measurement precision
If offline processing algorithms are used for landmark extraction, then extraction quality is improved, but processing time increases making it unsuitable for real-time autonomous vehicle applications
Solution Approach 1:
The patent applies dynamics by transitioning from static offline processing to dynamic real-time processing through adaptive model deployment. The system dynamically adjusts between different processing modes: using pre-trained lightweight models for real-time sparse point cloud processing during autonomous vehicle operation, while periodically updating models with new offline training data. This dynamic approach maintains high extraction quality while ensuring real-time performance capabilities.
Data Source
AI summary
A system and method for processing a 3D point cloud to generate a segmented point cloud in real time are disclosed, the method includes: receiving a sparse 3D point cloud captured by a detection and ranging sensor mounted to a vehicle, the 3D point cloud comprising a plurality of data points, each data point in the 3D point cloud having a set of coordinates in a coordinate system of the detection and ranging sensor; generating, from the 3D point cloud, a range map comprising a plurality of elements, each of the plurality of data points of the 3D point cloud occupying a respective element of the plurality of elements; labelling the data point in each respective element of the range map as one of a pole-like data point or a vertical-plane-like data point; and generating the segmented point cloud including one or more of the labeled data points.