Visual Localization Using Background-Weighted Feature Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The estimation of position and orientation using SLAM is compromised by moving subjects in photographic images, leading to reduced accuracy due to difficulties in collating images with environmental maps.
Innovation Solution
An information processing apparatus that calculates a weight for each feature point in an image based on its probability of being a stationary background point, using a combination of IMU measurements and deep neural networks, to enhance the accuracy of position and orientation estimation by prioritizing stationary features in the image processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If SLAM is used for position and orientation estimation by collating photographic images with environmental maps, then device localization can be achieved, but the presence of moving subjects in images deteriorates the estimation accuracy
Solution Approach 1:
The patent segments the image features by calculating individual weights for each feature point based on its probability of belonging to the stationary background. This segmentation allows the system to separate reliable background features from unreliable moving object features, thereby resolving the contradiction between achieving accurate localization and dealing with moving objects in the image.
Solution Approach 2:
The patent applies local quality by assigning different weights to different feature points based on their individual characteristics. Feature points with higher probabilities of being stationary background points receive higher weights, while those likely to be moving objects receive lower weights. This localized quality assessment improves overall estimation accuracy by emphasizing reliable features.
2Loss of information
If all feature points in the image are used for collation, then more information is available for estimation, but moving objects introduce errors that reduce accuracy
Solution Approach 1:
The patent implements partial action by selectively using only the reliable portion of image features for collation. By calculating weights and thresholding, the system uses a subset of feature points that are most likely to represent the stationary background, rather than using all feature points. This partial utilization of information maintains accuracy while avoiding the harmful effects of moving objects.
Solution Approach 2:
The system performs self-service by automatically evaluating the reliability of each feature point and adjusting its own collation process accordingly. The weight calculation and thresholding mechanisms enable the system to self-regulate which features to trust, eliminating the need for external intervention to filter out moving objects.
Data Source
AI summary
Provided is an information processing apparatus that includes a background weight calculation unit and a position/orientation estimation unit. The background weight calculation unit calculates, for each of a plurality of feature points included in a photographic image, a weight corresponding to a probability that the feature point is a point on a stationary subject forming a background of the photographic image. Based on a calculation result reflecting the weight of each of the feature points, the position/orientation estimation unit compares the photographic image with information regarding a keyframe image registered in an environmental map, and estimates position and orientation of the device that has captured the photographic image (IM).