Neural Image Localization for Drift-Free Trajectory Positioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional mapping and localization methods for moving objects in new environments suffer from errors due to visual feature selection issues and feature sparseness, leading to poor positioning accuracy.
Innovation Solution
A system using a neural network trained with reference images to predict the position of a moving object, anchored to exact positions, allowing for robust localization without drift and maintaining accuracy over long distances by employing dense image data and machine learning techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional mapping and localization methods are used, then the system can operate in new environments, but positioning accuracy deteriorates due to visual feature engineering errors and feature sparseness
Solution Approach 1:
The patent creates a database of reference images with known positions and uses a neural network to match current images against this database. Instead of building maps in real-time, the system copies and compares against pre-recorded reference data, achieving high accuracy by finding matches in the reference image database rather than constructing localization from sparse visual features in new environments
2Ease of operation
If visual features are used for localization, then the system can determine position, but accuracy deteriorates due to feature sparseness (only a few features present in each image)
Solution Approach 1:
The patent replaces traditional mechanical vision-based feature matching with a neural network that processes entire images. Instead of relying on sparse visual features and geometric calculations, the system uses deep learning to extract and match image representations, substituting the mechanical vision pipeline with an AI-based approach that achieves higher accuracy
3Measurement precision
If neural network with reference images is used, then positioning accuracy improves to a few centimeters, but device complexity increases
Solution Approach 1:
The patent performs preliminary actions by pre-recording reference images at known positions and pre-training the neural network on this data. This offline preparation creates a ready-to-use reference database and trained model, so that during actual operation the system only needs to perform image matching and position lookup, significantly reducing the computational complexity during runtime while maintaining high accuracy
Data Source
AI summary
Techniques for generating high-precision localization of a moving object on a trajectory are provided. In one technique, a particular image that is associated with a moving object is identified. A set of candidate images is selected from a plurality of images that were used to train a neural network. For each candidate image in the set of candidate images: (1) output from the neural network is generated based on inputting the particular image and said each candidate image to the neural network; (2) a predicted position of the particular image is determined based on the output and a position that is associated with said each candidate image; and (3) the predicted position is added to a set of predicted positions. The set of predicted positions is aggregated to generate an aggregated position for the particular image.