Self-Supervised Temporal Masking for SLAM Robustness

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current SLAM methods face performance drops or failures when dealing with dynamic environments, particularly when moving objects occupy a significant portion of the scene, due to limitations in masking techniques that either miss relevant static objects or incorrectly identify static ones as moving.

Innovation Solution

A self-supervised learning-based method that determines optimal temporal masking for SLAM algorithms, using a prediction model to decide whether to apply spatial masking for each image or class of objects based on performance evaluation metrics, thereby improving the accuracy and robustness of trajectory estimation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If motion detection-based masking is used to eliminate moving objects, then the SLAM algorithm can process static scenes effectively, but it fails when moving objects occupy a significant portion of the scene or when static objects are incorrectly identified as moving

Engineering Contradiction:
ImproveSLAM algorithm performanceVSAvoidtrajectory estimation accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies dynamics by transitioning from static masking rules to dynamic, adaptive masking decisions. The system evaluates each image frame independently and adjusts masking decisions based on real-time analysis of motion patterns and scene characteristics, allowing the masking strategy to adapt to changing environmental conditions rather than relying on fixed thresholds

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameters of the masking approach by moving from binary masking decisions to probabilistic masking. The system calculates a probability score for each pixel or region indicating its likelihood of being dynamic, and uses this continuous parameter to make informed masking decisions, allowing for more nuanced handling of ambiguous cases

Inventive Principle:
Principle #35Parameter changes

2Reliability

If spatial masking is applied to remove moving objects, then the SLAM algorithm can focus on static features, but it may incorrectly mask static objects or miss relevant dynamic information

Engineering Contradiction:
Improvemasking accuracyVSAvoidstatic object information
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent applies local quality by treating different regions of the image differently based on their individual characteristics. Instead of applying uniform masking rules, the system analyzes local motion patterns, depth information, and semantic content to determine the appropriate masking strategy for each region, preserving static objects that appear dynamic while removing truly moving objects

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent implements feedback mechanisms where the system continuously monitors the performance of masking decisions and adjusts its strategy accordingly. By evaluating the impact of masking on trajectory estimation and using this feedback to refine future masking decisions, the system learns from its own performance and improves accuracy over time

Inventive Principle:
Principle #23Feedback

3Measurement precision

If manual annotation is used to train masking models, then the model can be trained with precise ground truth, but the process is time-consuming and costly

Engineering Contradiction:
Improvemodel training accuracyVSAvoiddata annotation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies self-service by enabling the system to annotate its own training data automatically. The system uses unsupervised learning techniques to generate pseudo-labels for training data without human intervention, and employs self-supervised learning methods to train models that can learn from raw data distributions, eliminating the need for manual annotation while maintaining training quality

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent uses copying by creating synthetic training data that replicates real-world scenarios. The system generates artificial images and annotations that mimic complex dynamic scenes, allowing the model to learn from synthesized data that captures the essential characteristics of real environments without requiring actual annotated real-world data

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4191530A1Simultaneous localization and mapping method incorporating self-supervised temporal masking and machine learning model for generating such masking
Publication Date: 2023.06.07 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP4191530A1 patent drawingFigure 1~2
  • EP4191530A1 patent drawingFigure 3
  • EP4191530A1 patent drawingFigure 4

AI summary

A computer-implemented method for the automatic learning of a prediction model (402) configured to determine, for a given image sequence associated with a spatial masking sequence of said images, a temporal masking sequence consisting, for each image, of a decision to apply all or part of the spatial masking to the image, the prediction model (402) being trained from training data to determine a temporal masking sequence that optimizes a predetermined performance criterion of an algorithm for the simultaneous mapping and localization of a mobile device capable of acquiring said image sequence and locating the mobile device in its environment.