Rank Order Filtering for Texture-Rich Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Augmented Reality (AR) applications on mobile platforms face challenges in robustly tracking objects in real-time due to the correspondence problem, especially with texture-rich objects, leading to erroneous feature correspondences and tracking failures.
Innovation Solution
The method involves capturing images of objects, detecting their edges, projecting a model object onto subsequent images, determining gradient profiles, computing radii, and using rank order filtering to generate feature candidate points, thereby reducing erroneous correspondences and improving tracking robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If nearest neighbor association is used to match model edge to image edge pixels, then the matching process is simple and fast, but the number of erroneous correspondences increases significantly with texture-rich objects
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing the ranked order of gradient magnitudes for each model edge pixel before tracking begins. This pre-computed ranking information is then used during real-time tracking to quickly identify corresponding image edge pixels without performing complex comparisons, thus maintaining fast tracking speed while improving accuracy for texture-rich objects.
2Reliability
If a large number of edge pixels are considered for matching, then the probability of finding correct correspondences increases, but the computational complexity and processing time increase
Solution Approach 1:
The patent reduces processing complexity by pre-computing the ranked order of gradient magnitudes for all model edge pixels before tracking. This preliminary action stores the ordering information that would otherwise require complex real-time comparisons, allowing the system to quickly identify candidate correspondences during tracking without performing computationally intensive operations at runtime.
Solution Approach 2:
The patent changes the parameter used for matching from direct gradient magnitude comparison to ranked order comparison. By transforming the matching criterion into a rank-based approach, the system can efficiently evaluate multiple edge pixels while maintaining robustness against variations in lighting and texture, thus improving reliability without proportionally increasing complexity.
3Measurement precision
If gradient profile analysis is performed to improve matching accuracy, then the number of erroneous correspondences is reduced, but the processing time increases
Solution Approach 1:
The patent applies preliminary action by pre-computing the ranked order of gradient magnitudes for each model edge pixel before tracking begins. This pre-computed ranking information is then used during real-time tracking to quickly identify corresponding image edge pixels without performing complex comparisons, thus maintaining fast tracking speed while improving accuracy for texture-rich objects.
Solution Approach 2:
The patent transforms the matching approach by using rank order of gradient magnitudes rather than absolute gradient values. This parameter transformation allows the system to be invariant to lighting changes and focus on the relative structure of edge profiles, improving precision while keeping processing efficient through the pre-computed ranking.
Data Source
AI summary
A method of real-time tracking of an object includes capturing a first and a second image of the object. The object is detected in the first image and movement of the object is tracked between the images. Tracking of the object includes obtaining an initial pose of the camera; projecting an image of a model object onto the second image; determining a gradient profile of the second image from an edge point of the model object along a first direction that is normal to the edge of the model object; computing a radius on the gradient profile; determining a rank order of the peaks of the gradient profile along the radius; comparing the rank order with a predetermined rank order to generate a feature candidate point; and reducing a distance along the first direction between the feature candidate point and the edge point on the edge of the model object.