Image Tracking via B-Spline Particle Filter
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face challenges in accurately tracking objects with changing shapes that cannot be represented by simple translations, expansions, or rotations, requiring high computational loads and limited applications, especially in moving images.
Innovation Solution
An image processing device and method that represent the shape of an object as a linear combination of reference shapes using a particle filter, allowing for efficient tracking and visual flair addition by adjusting coefficients and predicting shape transitions based on a transition model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the Condensation algorithm is used to track objects with changing shapes, then tracking precision is improved, but computational load increases significantly
Solution Approach 1:
The object shape is segmented into multiple reference shapes that can be combined through linear combination. Instead of tracking one complex changing shape directly, the system divides it into simpler reference shapes (e.g., basic geometric forms) and reconstructs the object shape by combining these references with time-varying coefficients, reducing computational complexity while maintaining tracking precision
Solution Approach 2:
The system changes parameters by representing object shape not as a fixed template but as a linear combination of reference shapes with dynamic coefficients. This parameter transformation allows the shape to adapt to changes over time while maintaining a computationally efficient representation through a small number of basis shapes
2Measurement precision
If manual contour extraction is performed for each frame, then contour accuracy is improved, but processing time increases
Solution Approach 1:
Reference shapes are prepared and stored in advance before actual tracking begins. These pre-computed reference shapes serve as a foundation that eliminates the need for complex real-time contour extraction, allowing the system to achieve accurate contours through efficient linear combination of pre-prepared shapes
Solution Approach 2:
Instead of extracting contours directly from each image frame through computationally intensive manual steps, the system creates a simplified copy representation using linear combinations of reference shapes. This copied representation captures the essential contour information with much lower computational cost
3Adaptability or versatility
If the dynamical contour model is used to track moving objects, then shape adaptation is improved, but computational volume increases
Solution Approach 1:
A small set of universal reference shapes is created that can represent multiple different object shapes through linear combination. These reference shapes serve multiple functions by adapting to various object forms (e.g., different hand gestures, facial expressions) without requiring separate models for each shape, reducing computational volume while maintaining shape adaptation capability
Data Source
AI summary
First, the data of a moving image that is captured is read for each frame, and whether to start tracking is determined based on the presence or absence of a target object (S20, S22). An edge image of the image frame is created after it is determined to start tracking (S24). Meanwhile, a particle is distributed in a space of a coefficient set for each control point sequence when the control point sequence of a B-spline curve representing the shape of the target object is represented in a linear combination of control point sequence of a B-spline curve representing a plurality of reference shapes that are made available in advance (S26). A particle is also distributed in the space of a shape-space vector (S28), the likelihood of each particle is observed, and the probability density distribution is acquired (S30). A curve obtained by weighting parameters by the probability density distribution and then averaging the weighted parameters is created as a tracking result (S32).