Camera Pose Estimation via Warped Image Library
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing camera pose estimation methods in augmented reality applications are unreliable due to noisy image measurements and motion blur, especially when they lack a-priori knowledge of the scene or rely on computationally intensive pose-invariant feature descriptors.
Innovation Solution
A method that captures a model image at a canonical pose, generates warped images through known transformations, and creates an image library with associated poses, allowing for camera pose initialization and re-initialization without prior scene knowledge or pose-invariant descriptors, by updating and matching features in the image library.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pose-invariant feature descriptors are used for camera pose estimation, then the reliability of pose estimation is improved, but the computational complexity increases significantly
Solution Approach 1:
The patent pre-generates an image library containing warped versions of the background scene at various camera poses during an offline phase. This preliminary action stores the computational complexity in advance, allowing the online pose estimation to simply compare current images against pre-computed warped images, thus achieving reliable pose estimation without real-time computational burden.
2Loss of time
If a-priori knowledge of the background scene is used, then the pose recovery speed is improved, but the adaptability to different scenes deteriorates
Solution Approach 1:
The patent creates a universal image library that can adapt to different scenes by capturing a model image of the current background scene and generating warped versions from it. This universal approach allows the same algorithm to work across different scenes without requiring scene-specific pre-processing, achieving both fast pose recovery and scene adaptability.
3Speed
If continuous pose updates are performed using instantaneous image measurements, then the real-time tracking capability is improved, but the stability deteriorates due to noisy measurements
Solution Approach 1:
The patent pre-generates multiple warped images representing different camera poses and stores them in an image library. When performing pose estimation, the system compares the current image against this pre-prepared library, cushioning against the instability caused by noisy instantaneous measurements by having multiple reference poses available for comparison rather than relying on continuous incremental updates.
Data Source
AI summary
A method of camera pose estimation is provided that includes capturing a model image of a scene at a canonical camera pose, generating an image library from warped images of the model image and the model image, wherein each warped image is a transformation of the model image at a different pre-determined camera pose, capturing an image of the scene as a user moves the camera, reporting the current camera pose as a camera pose of the image when the image is acceptable, conditionally adding the first image to the image library when the first image is acceptable, and re-initializing the current camera pose to a camera pose selected from the image library when the first image is not acceptable.