Video SLAM Shape-Prior Learning for Drift-Resistant Mapping

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing SLAM technologies for robotic applications, such as augmented reality and autonomous vehicles, face challenges in accurately estimating scene structure and camera motion from images without relying on hand-crafted shape priors, and struggle with positional drift and loop closure issues.

Innovation Solution

A novel Convolutional Neural Network (CNN) architecture, SLAM-Net, utilizes Generative Adversarial Networks (GANs) for unsupervised learning of shape priors, enabling direct and dense SLAM with real-time depth map estimation and camera motion estimation, while avoiding photometric error assumptions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional SLAM methods use hand-crafted shape priors, then the system can operate without supervision, but the accuracy of structure estimation deteriorates

Engineering Contradiction:
Improvestructure estimation accuracyVSAvoidlearning system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs self-service by automatically learning shape priors from unlabeled video data through unsupervised learning. The neural network trains itself using only the input video sequences without requiring external supervision or hand-crafted priors, thereby improving structure estimation accuracy while avoiding the complexity of manual prior design

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The invention changes the parameter of shape prior representation from hand-crafted mathematical models to learned neural network parameters. This transformation allows the system to adapt shape priors dynamically from data, improving estimation accuracy while keeping the learning process automated and supervised-free

Inventive Principle:
Principle #35Parameter changes

2Reliability

If SLAM methods rely on photometric error optimization, then computation can be performed, but positional drift occurs over time

Engineering Contradiction:
Improveposition estimation stabilityVSAvoidposition accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The system implements feedback mechanisms through loop closure detection and continuous refinement of the 3D map. When the ROV returns to a previously visited location, the system detects this and uses the feedback to correct accumulated position errors, thereby maintaining position accuracy and reducing drift over time

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The SLAM system performs multiple functions simultaneously: it builds a 3D map, tracks camera motion, detects loop closures, and refines position estimates. This multi-functionality allows the system to cross-validate information from different sources (visual odometry, map consistency, loop closure) to maintain reliable position estimation without drift

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If dense SLAM reconstructs all pixels in an image, then scene structure estimation improves, but computational time increases

Engineering Contradiction:
Improvescene structure accuracyVSAvoidreal-time processing capability
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system segments the image into multiple regions and processes them in parallel using a neural network architecture designed for dense prediction. By dividing the dense SLAM computation into manageable segments that can be processed simultaneously, the system achieves high-accuracy pixel-level structure estimation while maintaining real-time processing capability through efficient parallel computation

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3834172B1System and method of operation for remotely operated vehicles for simultaneous localization and mapping
Publication Date: 2026.04.15 OCEAN INFINITY (PORTUGAL) SA
  • EP3834172B1 patent drawingFigure 1A~1B
  • EP3834172B1 patent drawingFigure 2A~2B
  • EP3834172B1 patent drawingFigure 3A

AI summary

The present invention provides systems and methods for simultaneous localization and mapping from video with adversarial shape prior learning in real-time. For example, an unsupervised direct and dense SLAM may learn a geometry prior from data. Given a video sequence, a depth map of a target frame, as well as the target frame and the camera motions between the target frame and all the remaining frames may be output. Further, by fusing a camera motion estimate with a positional sensor' s output, positional drift and the need for loop closure can be avoided.