Variational Autoencoder Road Feature Decay Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for 3D road geometry modeling and feature detection in autonomous vehicles are resource-intensive, time-consuming, and costly, and struggle with unreliable feature detection due to feature decay over time, which can impact autonomous driving accuracy and efficiency.

Innovation Solution

A method using a variational autoencoder network to learn an implicit transformation and interpolate feature-space decay from a single new image, allowing for the prediction of feature decay stages and improved detection accuracy under varying conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional methods for 3D road geometry modeling and feature detection are used, then measurement accuracy can be maintained, but resource consumption increases significantly and processing time is extended

Engineering Contradiction:
Improvefeature detection accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent replaces traditional mechanical measurement systems with human measurement and calculation with an automated computer vision system that uses image data and machine learning algorithms to detect features and model road geometry, significantly improving processing efficiency while maintaining measurement accuracy

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transforms the approach by changing from direct 3D measurement to 2D image analysis with computational modeling, using parameter transformations to derive three-dimensional road geometry information from two-dimensional image data through automated processing

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If feature detection systems are used to detect features in varying conditions, then navigation capability is enabled, but detection reliability deteriorates due to feature decay over time

Engineering Contradiction:
Improvedetection under varying conditionsVSAvoidfeature detection reliability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies preliminary action by pre-training the machine learning model with synthetic images that represent various stages of feature decay before deployment. This allows the system to anticipate and adapt to degraded features in advance, maintaining detection reliability across different time periods and environmental conditions

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements dynamics by creating a dynamic training dataset that simulates feature decay over time using variational autoencoders. The system learns to detect features at different decay stages, making the detection system adaptive and reliable across varying temporal and environmental conditions

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If more training data covering various decay stages is collected, then model accuracy improves, but data collection time and resources increase

Engineering Contradiction:
Improvemodel detection accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses copying by generating synthetic training images through variational autoencoders that simulate feature decay. Instead of collecting real images at every decay stage, the system creates realistic synthetic copies representing different decay levels, significantly reducing data collection time while maintaining model training effectiveness

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent employs cheap short-living objects by using computational algorithms to generate synthetic training data rather than investing time and resources in long-term field data collection. The synthetic images serve as disposable but effective training samples that cover the full range of decay scenarios

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Data Source

PatentEP3648003B1Method and apparatus for predicting feature decay using variational auto-encoders
Publication Date: 2024.01.17 HERE GLOBAL BV
  • EP3648003B1 patent drawingFigure 1
  • EP3648003B1 patent drawingFigure 2
  • EP3648003B1 patent drawingFigure 3

AI summary

An apparatus, method and computer program product are provided for predicting feature space decay using variational auto-encoder networks. Methods may include: receiving a first image of a road segment including a feature disposed along the road segment; applying a loss function to the feature of the first image; generating a revised image, where the revised image includes a weathered iteration of the feature; generating a predicted image using interpolation between the image and the revised image of a partially weathered iteration of the feature; receiving a user image, where the user image is received from a vehicle traveling along the road segment; correlating a feature in the user image to the partially weathered iteration of the feature in the predicted image; and establishing that the feature in the user image is the feature disposed along the road segment.