Camera Motion Estimation Using Vector Median Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing camera motion estimation methods often get stuck to dominant foreground motion, are computationally complex, and may not accurately represent camera motion due to nonlinear relations and non-uniform error distributions, leading to suboptimal results in video processing applications.
Innovation Solution
The method starts with a true smooth motion estimation to obtain a Motion Vector Field, applies vector median filtering, and determines optimal block size and reduction stages for camera motion estimation, ensuring more precise and robust global motion estimation by distinguishing between camera and dominant foreground motion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If iterative camera motion estimation methods are used to handle nonlinear relations, then measurement precision improves, but device complexity increases
Solution Approach 1:
The patent replaces complex iterative estimation methods with a simple non-iterative approach using vector median filtering. This disposable-like principle uses a straightforward algorithm that does not require multiple iterations or complex computational resources, yet achieves sufficient accuracy for camera motion estimation by filtering motion vectors to eliminate foreground dominance.
Solution Approach 2:
The patent substitutes the mechanical/iterative estimation process with a filtering-based approach. Instead of using iterative optimization algorithms that require multiple passes and complex computations, the invention applies vector median filtering to directly obtain camera motion parameters from motion vectors, replacing the complex system with a simpler filtering mechanism.
2Device complexity
If linear regression is used for parameter estimation, then device complexity decreases, but measurement precision deteriorates due to non-uniform error distribution
Solution Approach 1:
The patent changes the estimation parameter from mean-based (linear regression) to median-based (vector median filtering). This parameter change makes the estimation robust to non-uniform error distributions and foreground motion dominance, as the median is inherently more resistant to outliers and non-Gaussian errors compared to the mean.
Solution Approach 2:
The patent uses a simple non-iterative vector median filtering approach that requires minimal computational resources. This disposable-like method does not require complex iterative optimization or assumptions about error distributions, making it computationally efficient while maintaining robustness against non-uniform errors.
3Measurement precision
If block matching operation is used to obtain Motion Vector Field, then measurement precision improves, but productivity decreases due to computational complexity
Solution Approach 1:
The patent extracts only the essential camera motion parameters from the full Motion Vector Field using vector median filtering. Instead of processing all motion vectors through complex iterative methods, the invention extracts the dominant background motion component efficiently, separating it from foreground motion and obtaining camera parameters directly without exhaustive processing.
Solution Approach 2:
The patent employs a computationally inexpensive vector median filtering operation that can be applied directly to the Motion Vector Field. This disposable-like approach avoids the need for heavy iterative optimization, providing a fast and efficient method to estimate camera motion from block matching results without sacrificing accuracy.
4Productivity
If dominant motion is used as camera motion estimate, then productivity increases by simplifying the process, but measurement precision deteriorates due to foreground motion dominance
Solution Approach 1:
The patent applies local quality by treating different regions of the Motion Vector Field differently through vector median filtering. The filtering process identifies and suppresses local foreground motion dominance while preserving the global camera motion pattern, allowing the system to handle local variations in motion characteristics without compromising overall accuracy.
Solution Approach 2:
The patent introduces vector median filtering as an intermediary between the raw Motion Vector Field and the final camera motion estimate. This intermediary filtering step eliminates the direct use of dominant motion vectors that may be contaminated by foreground objects, providing a mediating process that separates true camera motion from spurious foreground motion while maintaining processing efficiency.
Data Source
Figure 1
Figure 2
AI summary
Camera motion estimation is required to obtain accurate global motion estimation (GME) for video applications. Thus, present invention describes a method for camera motion estimation by using an initial motion vector field which is obtained from smooth motion estimation unit and by using vector median filtering. In present invention the number of reduction stages according to initial motion vector field is determined and for each reduction stage, vector median filtering is applied to present reduction stage motion vector field.