Bias-Aware Training Data Creation for Road Element Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately classifying road elements due to bias in sensed information units, particularly under harsh conditions such as partial occlusions and varying illumination, which affects the reliability of machine learning processes.
Innovation Solution
A method involving the generation of artificially generated sensed information units to create a dataset that includes both biased and non-biased data, allowing the training of a machine learning process to treat each road element as a separate item of classification, thereby overcoming bias and improving classification accuracy under diverse conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning processes are trained on real sensed information units only, then the training data reflects actual road conditions, but bias occurs under harsh conditions such as partial occlusions and varying illumination
Solution Approach 1:
The patent creates artificial copies of sensed information units by generating synthetic images that replicate real road scenarios with controlled variations in occlusion, illumination, and road elements. These synthetic copies are then mixed with real data to train the machine learning model, enabling it to learn robust features that generalize across biased and unbiased conditions without requiring extensive manual data collection for each scenario
Solution Approach 2:
The system systematically varies parameters such as illumination levels, occlusion程度, weather conditions, and road element configurations in the synthetic data generation process. By changing these parameters across multiple dimensions, the training dataset covers a wide range of harsh conditions, enabling the model to maintain accurate classification across diverse and biased real-world scenarios
2Reliability
If more diverse training data is collected to cover harsh conditions, then classification accuracy under diverse conditions improves, but data collection and processing time increases
Solution Approach 1:
Instead of physically collecting diverse data under various harsh conditions, the system generates synthetic copies of road scenes programmatically. This approach creates unlimited diverse training samples instantaneously without the time-consuming process of capturing real-world data under each condition, while maintaining realistic visual characteristics that teach the model robust classification
Solution Approach 2:
The system performs preliminary generation of synthetic training data that anticipates future harsh conditions the vehicle may encounter. By pre-generating comprehensive coverage of occlusion, illumination, and weather variations before deployment, the model is already trained to handle these conditions when encountered in real operation, eliminating the need for real-time data collection and adaptive learning
3Reliability
If synthetic data is generated to overcome bias, then classification accuracy improves, but computational resources for data generation increase
Solution Approach 1:
The system generates synthetic copies of sensed information units using efficient rendering techniques that replicate the visual characteristics of real road scenes without requiring full-physics simulations. These lightweight synthetic copies provide sufficient training value while consuming minimal computational resources compared to complex simulation approaches
Solution Approach 2:
The synthetic data generation system is designed to serve multiple functions simultaneously: it generates training data for various road conditions, applies different occlusion and illumination scenarios, and creates augmented datasets all through a single integrated process. This multi-functionality reduces overall computational overhead by avoiding separate processing pipelines for each data augmentation task
Data Source
AI summary
A method of automatic bias related dataset creation for machine learning training, the method includes identifying, via a self-supervised learning process, a sensed information unit that is classification biased as it exhibits a combination of features, the sensed information unit is of a dataset associated with captured data in a road environment; automatically artificially creating a set of sensed information units exhibits only one or only some features of the combination of features; and adding the automatically artificially created set to the dataset to provide an updated data set in association with the identified classification biased sensed information unit for training a machine learning process with the updated dataset to provide a trained machine learning process that identifies each of the combination of features as a separate feature for classification.