Image Processing Device Scale Estimation via Discrete Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object tracking algorithms face challenges in estimating object scale accurately and in real-time, particularly when object motion is rapid or scale changes are large, leading to increased computational time and difficulties in real-time applications.
Innovation Solution
An image processing device that generates scaled samples from a region of interest, extracts features, and uses likelihood estimation and probability distribution learning to derive an estimated probability distribution, combining previous object features into a template for outlier rejection and feature matching to select the most similar sample as the final output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If scale estimation is performed using correlation filters with regression problem formulation, then scale estimation capability is improved, but computational time increases making real-time processing difficult
Solution Approach 1:
The patent segments the scale estimation problem into multiple discrete scale samples rather than treating it as a continuous regression problem. By dividing the scale space into discrete samples and using exhaustive search or candidate selection, the method avoids the computational burden of continuous optimization while maintaining accuracy.
Solution Approach 2:
The patent uses pre-computed scale samples that can be quickly evaluated without requiring complex iterative optimization. Each scale sample is a self-contained unit that can be assessed independently and discarded after evaluation, avoiding the need for persistent computational structures.
2Measurement precision
If the number of scale samples is increased to handle large scale changes, then scale estimation accuracy is improved, but the algorithm run time increases beyond real-time limits
Solution Approach 1:
The patent generates scale samples that may exceed the actual object scale (excessive action) but uses a scoring mechanism to identify the correct scale among them. This allows using a fixed, manageable number of samples regardless of the actual scale range, maintaining real-time performance while handling large scale variations.
Solution Approach 2:
The patent changes the parameter representation from continuous scale values to discrete scale samples with associated scores. By transforming the scale estimation into a selection problem among predefined samples rather than continuous optimization, the method achieves both accuracy and real-time performance.
3Device complexity
If correlation filters assume symmetric Gaussian distribution for scale estimation, then computational complexity is reduced, but accuracy deteriorates when scale changes are asymmetric
Solution Approach 1:
The patent explicitly handles asymmetric scale changes by treating scale estimation as a selection problem among discrete samples rather than assuming symmetric Gaussian distribution. Each scale sample is evaluated independently, allowing the method to capture asymmetric scale variations without requiring symmetric probability assumptions.
Data Source
AI summary
An image processing device according to one of the exemplary aspects of the present invention includes: a scale space generation means for generating the scaled samples from a given input region of interest; feature extraction means for extracting features from the scale samples; a likelihood estimation means for deriving an estimated probability distribution of the scaled samples by maximizing the likelihood of a given scaled sample and the parameters of the distribution; a probability distribution learning means for updating the model parameters given the correct distribution of the scaled samples; a template generation means to combine the previous estimates of the object features into a single template which represents the object appearance; an outlier rejection means to remove samples which have a probability below the threshold; and a feature matching means for obtaining the similarity between a given template and a scaled sample and selecting the sample with the maximum similarity as the final output.