Active Learning for Vehicle Object Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing active learning methods for automotive perception tasks in vehicles with partial driving automation lack effective criteria to ensure the accuracy required for these tasks, which is costly to achieve with large, labeled datasets.

Innovation Solution

A method that uses a latent space vector from a machine learning algorithm trained for object classification, calculates variance and distance quotients to determine the relevance of automotive sensor data for active learning, and provides this data to an object classifier to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If large, labeled datasets are used to train machine learning algorithms for automotive perception tasks, then the accuracy of object classification is improved, but the cost of data generation increases significantly

Engineering Contradiction:
Improveobject classification accuracyVSAvoidcost of data generation
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system performs self-labeling by using the machine learning algorithm to generate predictions that automatically serve as labels for training data, eliminating the need for expensive manual annotation while maintaining classification accuracy

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system changes the parameter of data labeling from manual human annotation to automated algorithmic prediction, fundamentally altering how training data is generated and reducing costs while preserving accuracy requirements

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If active learning criteria are used to select subsets of unlabeled data for labeling, then the cost of data generation is reduced, but the accuracy required for partial driving automation may not be achieved

Engineering Contradiction:
Improvecost of data generationVSAvoidobject classification accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The system uses feedback loops where the machine learning algorithm's predictions are continuously evaluated against actual outcomes, and this feedback informs the selection of subsequent data subsets for labeling, ensuring accuracy improvement while controlling costs

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system performs preliminary filtering and selection of unlabeled data subsets based on active learning criteria before they are labeled, prioritizing data that will most effectively improve model accuracy while minimizing labeling costs

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If manual labeling is performed to create large, labeled datasets, then the accuracy of machine learning algorithms is improved, but the time and resources required increase

Engineering Contradiction:
Improvemachine learning algorithm accuracyVSAvoidtime for data labeling
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system replaces the mechanical process of manual human labeling with an automated computational process using machine learning algorithms, dramatically reducing the time and labor required while maintaining or improving labeling consistency and accuracy

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

Data Source

PatentEP4567743A1Active learning for object classification in visual perception tasks in a vehicle
Publication Date: 2025.06.11 BAYERISCHE MOTOREN WERKE AG
  • EP4567743A1 patent drawingFigure 1
  • EP4567743A1 patent drawingFigure 2
  • EP4567743A1 patent drawingFigure 3

AI summary

A machine learning algorithm trained to classify objects in automotive sensor data is provided with automotive sensor data as an input. Based on this input, the machine learning algorithm determines an object class. Then, a latent space vector corresponding to activation values within a latent space of the machine learning algorithm is obtained and at least one latent space vector variance is calculated. If the at least one latent space vector variance exceeds a variance threshold, a latent space vector distance quotient is determined, which is indicative of a distance of the latent space vector to a closest object class latent space vector of the same object class relative to a distance of the latent space vector to a closest latent space vector of a different object class. Finally, the automotive sensor data is provided to an object classifier based on the latent space vector distance quotient.