Object Recognition via Keypoint Position Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition systems face challenges in reliably recognizing objects in input images when keypoints are insufficiently extracted due to conditions like occlusion, extreme size, or low resolution, leading to misjudgments about the presence of matching objects.
Innovation Solution
An object recognition apparatus that extracts typical feature points from a typical image, calculates their features, and compares them with object feature points from an input image, focusing on positional relationships to determine similarity, even when the number of keypoints is small, and recognizes objects based on these relationships and a predetermined ratio of corresponding points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If keypoints are extracted using conventional methods (DoG filter, extremal value detection), then scale-invariant and rotation-invariant features can be obtained, but the number of extracted keypoints becomes insufficient when objects are occluded, extremely small, or low-resolution
Solution Approach 1:
The patent segments the object detection task into two independent stages: first detecting corner points using conventional DoG methods, then detecting line points using gradient-based methods. This segmentation allows each method to optimize for its specific function, with corner points providing scale/rotation invariance and line points providing structural information, thereby compensating for the insufficiency of corner points alone.
Solution Approach 2:
The patent merges corner point detection and line point detection into a unified object recognition framework. By combining both detection results and using line points to supplement corner point information, the system achieves more reliable object recognition when corner points are insufficient, directly resolving the contradiction between recognition reliability and keypoint quantity.
2Adaptability or versatility
If only corner points are used for object recognition, then scale-invariant and rotation-invariant features can be obtained, but recognition becomes unreliable when objects are occluded or extremely small
Solution Approach 1:
The patent segments feature detection into corner points (for scale/rotation invariance) and line points (for structural information). This segmentation allows the system to maintain adaptability to scale and rotation changes while improving reliability through the complementary information provided by line points, especially in occluded or small object scenarios.
Solution Approach 2:
The patent applies different detection methods and quality criteria to different types of features. Corner points use DoG filtering with extremal value detection optimized for invariance, while line points use gradient-based detection optimized for structural information. This local quality differentiation allows each feature type to excel at its intended function, resolving the contradiction between invariance and reliability.
3Stability of the object's composition
If keypoints are removed based on contrast and edge criteria to extract stable keypoints, then keypoint stability improves, but the number of remaining keypoints decreases
Solution Approach 1:
The patent segments keypoint extraction into two parts: stable corner points (using contrast and edge filtering) and additional line points (using gradient detection). This segmentation allows the system to maintain keypoint stability through careful selection of corner points while compensating for their数量 deficiency with line points, thereby resolving the contradiction between stability and quantity.
Data Source
AI summary
An object recognition apparatus sets each of keypoints extracted from a typical image and one of keypoints extracted from an input image, having scale invariant features similar to each other, as a typical corresponding point and an object corresponding point matching with each other, and produces a position vector directed from a typical reference point of the typical image to each typical keypoint. The apparatus determines a position of an object reference point in the input image from a position of each object corresponding point and the vector of the typical corresponding point matching with the object corresponding point. When the positions of the object reference point are concentrated, the apparatus judges that an object picture having the object corresponding points in the input image matches with the typical image, and the apparatus recognizes the picture as an object indicated by the typical image.