Video Background Subtraction via Factorized Matrix Completion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing background subtraction methods fail to effectively segment moving objects from a video when the camera is moving, as they rely on low rank structure that no longer holds, and require complex singular value decompositions for large dimensions.
Innovation Solution
The method employs factorized robust matrix completion (FRMC) with global motion compensation, decomposing video images into a low rank background and sparse motion components using a Pareto curve approach, and aligns images using motion vectors to account for camera perspective changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If algebraic decomposition with low rank structure is used for background subtraction, then segmentation accuracy is improved for stationary cameras, but the method fails when the camera is moving because the low rank structure no longer holds
Solution Approach 1:
The patent applies dynamics by transitioning from a static low-rank background model to a dynamic model that incorporates motion vectors. The background is represented as a low-rank matrix that is updated over time based on estimated motion, allowing the system to adapt to camera movement while maintaining the computational efficiency of low-rank decomposition. This resolves the contradiction by making the background model dynamic rather than static.
Solution Approach 2:
The patent changes the parameters of the background model by introducing motion vectors as additional variables. Instead of assuming a fixed low-rank structure, the system estimates motion parameters from the video data and uses these to update the background representation. This parameter change enables the system to handle moving cameras while maintaining segmentation accuracy.
2Measurement precision
If robust principal component analysis (RPCA) is used for background subtraction, then foreground segmentation is achieved, but complex singular value decompositions are required for large dimensions making real-time processing difficult
Solution Approach 1:
The patent segments the background subtraction problem into two independent subproblems: one for the low-rank background component and one for the sparse foreground component. By decomposing the original large-scale RPCA problem into these smaller, manageable subproblems, the system reduces computational complexity while maintaining segmentation accuracy, enabling real-time processing.
Solution Approach 2:
The patent applies partial action by using truncated singular value decomposition instead of full SVD. Instead of decomposing the entire large matrix, the system only computes the top k singular values and vectors, where k is much smaller than the actual rank. This partial decomposition significantly reduces computational cost while retaining the essential information needed for accurate foreground segmentation.
3Reliability
If full singular value decomposition is performed for large dimensional video data, then complete background modeling is achieved, but computational complexity increases making real-time processing infeasible
Solution Approach 1:
The patent applies partial action by using truncated singular value decomposition that computes only the top k singular values and vectors needed for accurate background modeling. This partial decomposition maintains reliability by capturing the essential low-rank structure while dramatically reducing computational complexity compared to full SVD, making real-time processing feasible.
Solution Approach 2:
The patent changes the approach by parameterizing the background as a low-rank matrix with rank k, where k is much smaller than the actual dimensionality of the video data. This parameter change allows the system to achieve accurate background modeling without performing computationally intensive full-rank decompositions, thus reducing device complexity.
Data Source
AI summary
A method processes a video acquired of a scene by first aligning a group of video images using compressed domain motion information and then solving for a low rank component and a sparse component of the video. A homography map is computed from the motion information to determine image alignment parameters. The video images are then warped using the homography map to share a similar camera perspective. A Newton root step is followed to traverse separately Pareto curves of each low rank component and sparse component. The solving for the low rank component and the sparse component is repeated alternately until a termination condition is reached. Then, the low ranks component and the sparse component are outputted. The low rank component represents a background in the video, and the sparse component represents moving objects in the video.