Computer Vision Localization Using Semantic Landmarks Without GNSS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing navigation systems, particularly for unmanned vehicles, face challenges in determining position when satellite signals are unavailable or unreliable, and conventional image-based feature mapping requires large datasets and lacks distinctive features for accurate positioning over wide areas.
Innovation Solution
A computer vision-based method using Deep Convolutional Neural Networks (CNN) and Region Proposals to identify landmarks, combined with a Monte Carlo Localization algorithm, allows for position determination and navigation without relying on GNSS, by classifying objects and landmarks through a semantic object dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image-based feature mapping is used to determine position, then location identification can be achieved, but it requires access to an unrealistically large dataset of image information and fails to provide distinctive features for accurate positioning over wide areas
Solution Approach 1:
The patent extracts and utilizes only the distinctive geometric features of landmarks (shape, size, orientation, spatial relationships) rather than requiring complete image datasets. This extraction approach reduces the data burden while maintaining positioning accuracy by focusing on key identifying characteristics of landmarks.
Solution Approach 2:
The system performs preliminary identification and classification of landmarks by type (building, tree, road, signpost, etc.) with their geometric properties stored in advance. This preliminary action enables the system to quickly match observed features with database entries without requiring large-scale image processing during real-time navigation.
2Reliability
If satellite signals are used for navigation, then position determination can be achieved, but it becomes unavailable or unreliable in underground areas, multi-storey structures, or where satellite coverage is inadequate
Solution Approach 1:
The patent creates a universal navigation system that functions across diverse environments by replacing satellite-dependent methods with landmark-based computer vision. The system can operate in underground areas, multi-storey structures, and open spaces alike by identifying and matching geometric features of landmarks in the environment, achieving environment-independent navigation capability.
3Ease of operation
If conventional pattern recognition functions are used to identify landmarks, then basic feature detection can be achieved, but the identified features are not sufficiently distinctive to enable positions to be determined over a wide area
Solution Approach 1:
The patent enhances feature distinctiveness by incorporating multiple dimensions of landmark characteristics: geometric shape, size, orientation, spatial relationships between features, and semantic classification (building, tree, road, etc.). This multi-dimensional feature representation transforms simple pattern recognition into precise position determination capability.
Data Source
Figure 1~2
Figure 3
Figure 4A~5
AI summary
In relation to the field of vehicle navigation, we describe a method of determining a position of a subject, comprising the steps of obtaining and storing an object dataset comprising object data indicative of multiple objects in an environment, including an indication of object parameters associated with each object, the object parameters including the location of the object in the environment and a semantic type classification associated with the object, obtaining environment data indicative of a region of the environment from a sensor associated with the subject, determining the presence of multiple observed objects in the environment data using an Artificial Neural Network to detect and classify each observed object, including determining one or more equivalent observed object parameters associated with each observed object including its respective semantic type classification, and determining the position of the subject using a probabilistic model, based on a comparison of the observed object parameters of each of the multiple observed objects with the equivalent object parameters of the objects in the object dataset including a comparison of the respective semantic type classifications.