Deep Learning Position Estimation Using Multi-Resolution Sensor Grids

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing positioning technologies face a trade-off between real-time processing and accuracy in environments with many obstacles, with signal processing-based methods being inaccurate and machine learning-based methods being computationally complex and less efficient.

Innovation Solution

A preprocessing method involving dividing regions into coarse, fine, and finer grids to generate training data for a deep learning model, using similarity and difference values between virtual and real sensor measurements, and training multiple deep neural networks to estimate the position of a target object.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If accurate preprocessing of wireless signal data is performed to enhance positioning accuracy, then positioning accuracy is improved, but calculation complexity increases and real-time efficiency cannot be secured

Engineering Contradiction:

Improvepositioning accuracyVSAvoidcalculation complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the positioning process into multiple stages with different grid resolutions. First, a coarse grid is used for initial positioning to reduce search space. Then, progressively finer grids are applied in subsequent stages to refine the position estimate. This segmentation allows the system to achieve high positioning accuracy without requiring intensive calculation across the entire area at full resolution, thus resolving the contradiction between accuracy and computational complexity.

Inventive Principle:

Principle #1Segmentation

2Productivity

If preprocessing is minimized for real-time positioning, then real-time efficiency is improved, but positioning accuracy is reduced

Engineering Contradiction:

Improvereal-time efficiencyVSAvoidpositioning accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent implements a dynamic preprocessing approach where the grid resolution is adaptively adjusted based on the positioning stage. In early stages, a coarser grid with lower computational requirements is used to maintain real-time efficiency. As the positioning converges and higher accuracy is needed, the system dynamically transitions to finer grid resolutions. This dynamic adjustment allows the system to optimize the balance between real-time efficiency and positioning accuracy throughout the processing sequence.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS12578418B2Preprocessing method of generating training data of deep learning model for estimating position of target object in environment having many obstacles, learning method of the deep learning model, and computing apparatus for estimating position of the target object

Publication Date: 2026.03.17 ELECTRONICS & TELECOMM RES INST

AI summary

A preprocessing method of a data preprocessing device is provided. The preprocessing method includes dividing a region, where the target object is located, into coarse grids and generating first sample data including a similarity value between a virtual sensor measurement value and a real sensor measurement value measured in each of the coarse grids, dividing each coarse grid into fine grids and generating second sample data including a similarity value between a virtual sensor measurement value and a real sensor measurement value measured in each of the fine grids, dividing each fine grid into finer grids and generating third sample data including a similarity value between a virtual sensor measurement value and a real sensor measurement value measured in each of the finer grids, and inputting the training data, including the first to third sample data, to the deep learning model.