Spatial Map Generation With Embedding Grids and Location Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional map generation methods for moving robots using TOF cameras suffer from reduced accuracy due to errors in location information, require high memory and processing power for high-resolution image rendering, and struggle with visual information capacity, leading to slow data processing and memory demands.
Innovation Solution
A map generation model utilizing an encoder and decoder module within a deep neural network to process captured images, generating spatial map data by embedding RGB and depth information into grids, and correcting location information through a loss function and location estimation/correction modules.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional grid-based map generation method is used to record three-dimensional distance information, then map structure information can be obtained, but map accuracy deteriorates as robot travel distance increases due to location information errors
Solution Approach 1:
The patent changes the parameter representation from simple occupancy grids to embedding vectors that encode multiple features (distance, color, texture, shape) in a continuous vector space. This transformation allows the system to maintain accurate spatial representations over long distances by incorporating relative position information and visual features that are invariant to cumulative drift errors.
Solution Approach 2:
The patent implements feedback mechanisms through loss functions that compare rendered images with captured images, and through location correction modules that use the generated map to correct accumulated location errors. This feedback loop continuously refines both the map accuracy and location information, preventing degradation over long travel distances.
2Measurement precision
If high-resolution image rendering is performed on two-dimensional or three-dimensional maps, then visual information quality is improved, but data processing speed decreases and memory requirements increase
Solution Approach 1:
The patent transitions from rendering high-resolution images on 2D/3D map surfaces to storing embedding vectors in a multidimensional vector space. Each grid cell contains a compact embedding vector that encodes visual information efficiently, eliminating the need for memory-intensive image rendering while preserving essential visual features for navigation and recognition.
Solution Approach 2:
Instead of storing and processing actual high-resolution images, the patent uses embedding vectors as compressed representations that capture essential visual features. These embeddings serve as efficient proxies for full images, enabling fast processing and low memory usage while maintaining the ability to reconstruct or recognize visual content when needed.
3Quantity of substance
If conventional methods store only entire structure and obstacle presence at high density, then memory usage is reduced, but visual information such as object shapes is lost and expressed only as sparse graphs
Solution Approach 1:
The patent merges multiple types of information (occupancy, distance, color, texture, shape) into unified embedding vectors that are stored in each grid cell. This integration allows the system to maintain compact memory usage while preserving rich visual information, as all these features are encoded together in a single vector representation rather than requiring separate data structures.
Solution Approach 2:
The patent transforms visual information from discrete graphical representations (sparse graphs) to continuous embedding vectors that can encode nuanced visual features including object shapes, colors, and textures. This parameter transformation enables efficient storage while maintaining detailed visual information that was previously lost in conventional sparse representations.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enhances map accuracy and reduces processing time by directly recording embedding data in grids, allowing for high-density image capture and efficient location estimation and correction, enabling precise navigation.
Implementation Method 1
generates embedding data by applying captured images taken by a movement device moving in a training space to an encoder module
Implementation Method 2
a time of flight (TOF) camera may be used. The TOF camera is a device that obtains three-dimensional distance information by using a TOF method for measuring the time when infrared rays emitted from a light emitting unit are reflected by an object and return to a light receiving unit
Data Source
AI summary
A map generation model building device includes a memory storing a map generation model building program, and a processor configured to execute the program, wherein the program generates embedding data by applying captured images taken by a movement device to an encoder module, generates spatial map data by recording the embedding data in map base data based on location information of the movement device, generates a rendering image based on the location information of the movement device in the spatial map data by using a decoder module, and train the a map generation model by comparing the rendering image with the captured image through a loss function and by updating the encoder module and the decoder module, the map base data includes a plurality of grids in which the embedding data is recorded, the embedding data includes RGB information and depth information for each pixel of the captured image.