Ambiguous Lane Event Mining for Autonomous Model Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in reacting to rare events due to the complexity and unpredictability of road and traffic conditions, and existing data collection methods are inefficient and costly, particularly for lane marker detection.
Innovation Solution
A system and method for lane marker detection that identifies ambiguous images, labels them, and adds them to a training corpus for retraining autonomous driving models, using a computer system to determine pixel fractions and flag interesting events for targeted data collection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large amounts of data from individual vehicles are transmitted to remote servers for storage and analysis, then the autonomous driving model can learn from existing road and traffic conditions, but valuable communication bandwidth is consumed and costs become prohibitively expensive
Solution Approach 1:
The patent segments the data transmission process by filtering and selecting only ambiguous lane detection events for transmission to remote servers, while processing and storing clear events locally. This segmentation reduces the volume of transmitted data significantly, addressing the bandwidth consumption problem while still providing sufficient training data for model improvement.
Solution Approach 2:
The patent implements local quality assessment by evaluating each lane detection event's ambiguity level at the vehicle端 using confidence scores from the autonomous driving model. Events are classified as ambiguous or clear based on local analysis, enabling selective transmission of only those events that require remote processing and model retraining.
2Quantity of substance
If all captured road condition data is collected and stored for model training, then the training dataset is comprehensive, but the data volume is excessively large and costs increase
Solution Approach 1:
The patent applies partial action by transmitting only a subset of captured events—specifically, ambiguous lane detection events that fall within predefined confidence score thresholds—to remote servers for storage and model retraining. This selective approach provides sufficient training data for model improvement while significantly reducing data transmission and storage costs compared to transmitting all captured events.
3Adaptability or versatility
If the autonomous driving model is trained on rare and ambiguous events, then the model's ability to handle edge cases improves, but the complexity of data processing and model retraining increases
Solution Approach 1:
The patent implements preliminary action by pre-processing and filtering lane detection events at the vehicle端 before transmission. Ambiguous events are identified and selected based on confidence score thresholds, and only these pre-filtered events are transmitted to remote servers for model retraining. This preliminary filtering reduces the complexity of remote data processing while ensuring that the most valuable training data is prioritized.
4Reliability
If ambiguous lane detection events are identified and transmitted for retraining, then the model learns from challenging cases, but the time required for event identification, labeling, and retraining increases
Solution Approach 1:
The patent implements self-service by enabling autonomous vehicles to automatically identify, classify, and transmit ambiguous lane detection events without requiring manual intervention. The system uses predefined confidence score thresholds to automatically filter and prioritize events, reducing the time and resources needed for manual data processing and preparation for model retraining.
Data Source
AI summary
A computer system obtains a road image captured by a vehicle. The computer system determines whether the road image is ambiguous for lane marker classification. When the computer system determines that the road image is ambiguous for lane marker classification, the computer system generates a labeled road image and adds the labeled road image to a corpus of training data for training a model to generate an autonomous driving model. The computer system distributes the autonomous driving model to one or more vehicles. The autonomous driving model is configured to process road images captured by the one or more vehicles to facilitate at least partially autonomously driving the one or more vehicles.