Training Frame Selection Using Uncertainty and Diversity Sampling

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing active learning processes for autonomous vehicles often select training frames from a single high-uncertainty traffic situation, neglecting frames from other situations with similar uncertainty, leading to incomplete model training and high recognition uncertainty for diverse scenarios.

Innovation Solution

A method that selects training frames with high recognition uncertainty and applies diversity sampling to ensure frames from different traffic situations are included, using fingerprint data and diversity criteria to maximize model improvement.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If frames are selected based solely on highest recognition uncertainty, then the number of annotated training examples is reduced, but the model training becomes incomplete and recognition uncertainty remains high for diverse scenarios

Engineering Contradiction:

Improvetraining efficiencyVSAvoidmodel training completeness

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent changes the selection criterion from a single parameter (recognition uncertainty) to multiple parameters by introducing fingerprint data that characterizes traffic situations. This allows the system to select frames based on both uncertainty levels and situational diversity, resolving the contradiction between training efficiency and completeness

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent combines multiple selection criteria (recognition uncertainty + fingerprint data diversity) into a composite selection mechanism. This composite approach ensures that selected frames both improve the model (high uncertainty) and cover diverse scenarios (fingerprint diversity), simultaneously achieving training efficiency and completeness

Inventive Principle:

Principle #40Composite materials

2Ease of operation

If only frames from a single traffic situation are selected, then the annotation process is simplified, but the model exhibits high recognition uncertainty for other traffic situations

Engineering Contradiction:

Improveannotation process simplicityVSAvoidmodel performance across diverse scenarios

Core Design Contradiction:

Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent performs preliminary analysis by calculating fingerprint data for all frames before selection. This preliminary characterization of traffic situations enables the subsequent selection process to efficiently choose diverse frames without complex real-time analysis during annotation, maintaining operational simplicity while ensuring versatility

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentEP4356353B1Method and electronic processor circuit for selecting a set of training frames from a plurality of given frames in an active learning process and computer-readable storage medium

Publication Date: 2026.02.18 CARIAD SE

Figure 1
Figure 2

AI summary

The invention is concerned with a method for selecting a set of training frames (24) from a plurality of given frames (19) of a frame sequence (18) in an active learning process (21), comprising the steps of providing a pre-trained machine learning model (22) and performing a respective recognition step (36) for the given frames (19) using the pre-trained model (22) and thereby generating a respective recognition result (37). In a pre-selection step (38) a respective uncertainty value (40) for each given frame (19) is calculated and N given frames (19) are chosen whose uncertainty values (40) indicate the highest recognition uncertainty. The invention is characterized by the selection step (42) of selecting the set of K training frames (24) from the N pre-selected frames (41) by performing a predefined diversity sampling.