Object Tracking via Singular Point Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for tracking objects in image sequences, such as optical flow and point matching, struggle when objects have large displacements or repetitive patterns, leading to ambiguity and errors in estimating the object's displacement model.
Innovation Solution
A method that selects singular points from the previous image based on low dissimilarity thresholds and ordered lists of points, forming pairs with the current image, and estimates the displacement model using a majority vote from these selected pairs, incorporating geometric constraints to reduce ambiguity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optical flow method is used to track object displacement, then tracking accuracy is improved for small displacements, but reliability deteriorates when displacement amplitude is large or unknown
Solution Approach 1:
The patent changes the fundamental parameters of the tracking approach by switching from global optical flow methods to local point matching methods. This involves extracting local descriptors (SIFT, SURF) at characteristic points and computing similarity metrics, which allows the system to adapt to various displacement amplitudes without relying on small displacement assumptions.
Solution Approach 2:
The patent segments the object tracking problem into local point matching tasks rather than solving it globally. By extracting characteristic points and computing local descriptors at discrete locations, the system can handle large displacements at each local point independently, improving overall reliability while maintaining accuracy.
2Reliability
If point matching method is used to handle large displacements, then reliability is improved, but measurement precision deteriorates due to ambiguity in matching repetitive patterns
Solution Approach 1:
The patent introduces asymmetry in the matching process by evaluating not only the similarity between point pairs but also the geometric consistency and spatial relationships. By considering the arrangement of multiple characteristic points and their relative positions, the system can disambiguate between multiple matching candidates that have similar local descriptors, thereby improving precision.
Solution Approach 2:
The patent implements feedback mechanisms through iterative refinement processes. The system initially matches points based on descriptor similarity, then uses geometric constraints and consistency checks to validate and correct mismatches. This feedback loop allows the system to resolve ambiguities in repetitive patterns and improve the precision of displacement estimation.
3Productivity
If all matched points are used to estimate displacement model, then productivity is improved, but measurement precision deteriorates due to inclusion of erroneous matches
Solution Approach 1:
The patent applies local quality by evaluating and weighting different matched points differently based on their reliability. Not all matched points are treated equally; instead, the system assigns higher weights to matches that satisfy geometric constraints and consistency checks, while reducing the influence of questionable matches. This selective approach maintains productivity while improving precision.
Solution Approach 2:
The patent uses partial action by selecting a subset of high-quality matched points for final displacement model estimation, rather than using all extracted points. By filtering and selecting only the most reliable matches that satisfy multiple criteria, the system achieves more accurate results without significantly reducing processing efficiency.
Data Source
Figure 1~2
Figure 3~5
Figure 4
AI summary
The present invention relates to a method of following an object in a sequence of at least two images termed previous and current. The said method comprises a step for forming a first set Ep of points Ep = {Pp (1),...,Pp (i),...,Pp (N)} by extracting N characteristic points Pp (i) of the object present in the previous image and for forming a second set Ec of points Ec = {Pc(1),...,Pc (i),..., Pc (M)} by extracting M characteristic points Pc(j) of the object present in the current image, a step for estimating the parameters of a model of displacement of the object between the two images on the basis of pairs of matched points thus formed, and a step of selecting the pairs of matched points used to estimate the parameters of the displacement model, in the course of which said pairs of matched points are selected solely from among those which are related to points of the first set of points which are singular, each point of the first set of points being a singular point.