3D Point Cloud Extraction Using Multiple Surface Orientation Hypotheses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional structure-from-motion algorithms have difficulty identifying vertical features, such as telephone poles, due to their inability to effectively match surface orientations in 2D imagery.
Innovation Solution
The system resamples local regions from non-reference images into a reference image's space multiple times with different surface orientation hypotheses, runs cross-correlation style correlators, and selects the hypothesis with the highest correlation score to determine a 3D location, allowing for more accurate mapping of vertical features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional structure-from-motion algorithms use cross-correlation style tiepoint generators to resample imagery into a common projection, then the algorithms can effectively identify features whose surface orientation matches the terrain (such as ground or building tops), but the algorithms have difficulty identifying vertical features (such as telephone poles)
Solution Approach 1:
The algorithm segments the surface orientation space into multiple discrete hypotheses (e.g., horizontal ground, vertical walls, sloped surfaces). Instead of assuming a single terrain model, the system divides the problem into multiple possible surface orientations and tests each hypothesis separately, allowing accurate identification of features with different orientations including vertical structures like telephone poles
Solution Approach 2:
The system dynamically adapts the surface orientation hypothesis based on the local terrain context. For each image ray and local region, the algorithm selects appropriate surface orientation hypotheses from a set of possibilities, allowing the matching process to adapt to different terrain types (flat ground, slopes, vertical walls) rather than using a fixed terrain model
2Measurement precision
If the system resamples local regions multiple times with different surface orientation hypotheses and runs cross-correlation style correlators, then the system can accurately extract 3D points for vertical features, but the computational complexity increases
Solution Approach 1:
The system performs resampling and correlation testing for only the necessary number of surface orientation hypotheses based on local terrain context, rather than exhaustively testing all possible orientations. By selecting a limited set of relevant hypotheses (e.g., horizontal, vertical, and a few intermediate orientations), the system achieves accurate 3D point extraction for vertical features while keeping computational complexity manageable
Solution Approach 2:
The algorithm performs preliminary classification of local terrain regions to identify which surface orientation hypotheses are relevant before performing the computationally intensive resampling and correlation operations. This preliminary action filters out irrelevant hypotheses, reducing the number of resampling operations needed and lowering overall computational complexity
Data Source
AI summary
The system and methods described herein operate on a plurality of images that include multiple views of the same scene, typically from slightly different viewing angles and/or lighting conditions. One of the images is selected as a reference image. For each image ray in a non-reference image, the system and methods resample a local region from the non-reference image's space to the reference image's space. The resampling is performed multiple times, each time with a different surface orientation hypothesis. The system and methods run cross-correlation style correlators on the resampled images, evaluate correlation scores for each of the resampled images, and select the surface orientation hypothesis associated with the highest correlation score. The system and methods project a peak of the correlation surface back through a geometry model for the selected surface orientation hypothesis to determine a three-dimensional (ground) location for the image ray.