Lane Ambiguity Mining for Efficient Autonomous Model Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in training for rare and unpredictable road conditions due to the high cost and inefficiency of collecting and transmitting large amounts of data, particularly for ambiguous lane marker detection, which hampers their ability to react to unusual events.
Innovation Solution
A system and method that utilize an event miner to identify ambiguous lane marker classifications in road images, label them, and add them to a training corpus for retraining autonomous driving models, thereby improving the recognition of lane markers in ambiguous images and reducing data volume and transmission costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If large amounts of road image data are collected and transmitted from vehicles to remote servers for analysis, then the training dataset can be expanded to improve autonomous driving model performance, but communication bandwidth consumption increases and transmission costs become prohibitively expensive
Solution Approach 1:
The patent applies local quality by performing different processing operations on different portions of data based on their characteristics. Specifically, road images are analyzed locally in vehicles using on-board processors to identify ambiguous lane marker cases, while only these specific ambiguous cases are transmitted to remote servers for further processing. This selective approach expands the training dataset with high-value ambiguous cases without transmitting all collected road images, thereby reducing communication bandwidth consumption while improving model performance.
2Loss of information
If all collected road image data is transmitted to remote servers for storage and analysis, then complete data can be processed for model training, but transmission costs become prohibitively expensive and bandwidth consumption increases
Solution Approach 1:
The patent extracts only the most valuable and informative data portions for transmission. By using on-board processors to analyze road images and identify ambiguous lane marker cases, the system extracts specifically those images that will most benefit model training. This extraction approach ensures that the training dataset receives high-quality ambiguous cases while avoiding transmission of redundant or clearly classifiable images, thus reducing transmission costs while maintaining data completeness for training purposes.
3Loss of energy
If on-board processors identify and flag only ambiguous lane marker cases for transmission, then bandwidth and cost are reduced, but the system complexity increases due to additional processing requirements
Solution Approach 1:
The patent applies preliminary action by performing initial analysis and filtering of road images using on-board processors before transmission. The on-board processors pre-identify ambiguous lane marker cases and flag them for transmission to remote servers. This preliminary processing step reduces the volume of data requiring transmission and simplifies server-side processing, as only pre-identified ambiguous cases need further analysis. The complexity added at the vehicle level is offset by the reduction in transmission requirements and server processing load.
Data Source
AI summary
A computer system obtains a plurality of road images captured by one or more cameras attached to one or more vehicles. The one or more vehicles execute a model that facilitates driving of the one or more vehicles. For each road image of the plurality of road images, the computer system determines, in the road image, a fraction of pixels having an ambiguous lane marker classification. Based on the fraction of pixels, the computer system determines whether the road image is an ambiguous image for lane marker classification. In accordance with a determination that the road image is an ambiguous image for lane marker classification, the computer system enables labeling of the image and adds the labeled image into a corpus of training images for retraining the model.