Keypoint Resampling Fusion for Visual Odometry Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI-based feature extraction methods in computer vision face limitations such as non-discriminative keypoint selection due to binary thresholding, leading to incorrect feature matching and spatially agonistic descriptors, which reduce the accuracy of pose estimation in applications like visual odometry and SLAM.
Innovation Solution
The proposed method employs Keypoint Resampling and Fusion (KRF) to generate a final set of keypoint-descriptor pairs by using multiple confidence thresholds, resulting in pixel-stacked high-confident keypoints and complementary descriptors, which improves the ratio of correct to incorrect correspondences and maintains local keypoint density.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If binary thresholding is used for keypoint selection, then the feature extraction process is simple and fast, but the keypoint selection becomes non-discriminative leading to incorrect feature matching
Solution Approach 1:
The patent changes the parameter from binary thresholding to multi-level confidence thresholding. Instead of using a single binary threshold, the system uses multiple confidence thresholds (first threshold and second threshold) to select keypoints at different confidence levels, thereby improving discriminative power while maintaining computational efficiency.
Solution Approach 2:
The patent segments the keypoint selection process into two distinct stages: first selecting keypoints above a higher confidence threshold, then selecting additional keypoints above a lower confidence threshold. This segmentation allows the system to prioritize high-quality keypoints while still incorporating lower-quality ones, improving overall matching accuracy without sacrificing speed.
2Reliability
If multiple keypoints are selected from the same position, then the number of high-quality keypoints increases, but the descriptor set becomes larger and more complex
Solution Approach 1:
The patent applies local quality by creating complementary descriptors for stacked keypoints at the same position. Each keypoint in the stack receives a descriptor with a unique quality indicator that reflects its confidence level, allowing the system to distinguish between high and low quality descriptors during matching while maintaining a manageable descriptor set size.
Solution Approach 2:
The patent introduces an intermediary mechanism (complementary descriptor encoding) that mediates between multiple keypoints at the same position. The complementary descriptor acts as an intermediary representation that encodes both the spatial position and the confidence level, enabling efficient matching without requiring separate full descriptors for each keypoint.
3Measurement precision
If high confidence threshold is used for keypoint selection, then the quality of selected keypoints is high, but the number of detected keypoints decreases
Solution Approach 1:
The patent introduces dynamics by implementing a two-stage keypoint selection process that adaptively selects keypoints based on confidence thresholds. The system dynamically adjusts the selection criteria: first selecting keypoints with high confidence above a higher threshold, then supplementing with additional keypoints above a lower threshold, thereby maintaining both quality and quantity.
Solution Approach 2:
The patent applies partial action by selectively applying different selection criteria to different batches of keypoints. The first batch is selected with strict high-confidence criteria to ensure quality, while the second batch uses more lenient criteria to increase quantity, achieving a balanced approach that satisfies both requirements.
Data Source
AI summary
A method of feature extraction from an image can include receiving the image including pixels, generating confidence values corresponding to positions of the pixels in the image by an artificial intelligence (AI) based feature extractor, selecting a first position among the positions of the pixels in the image, a first confidence value among the generated confidence values at the first position being higher than a first threshold, and generating a final set of keypoint-descriptor pairs based on the confidence values corresponding to positions of the pixels in the image. The final set of keypoint-descriptor pairs includes at least two keypoint-descriptor pairs corresponding to the first position among the positions of the pixels in the image.