Driving Scenario Discovery for Rare-Event Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning models for vehicle driving systems struggle with predictive accuracy for scenarios they have not previously encountered, particularly for rare or critical driving conditions such as slippery roads or emergency vehicles.
Innovation Solution
A processor-readable medium and method for natural language-based scenario discovery, which involves querying image data, generating scores, and iteratively retraining classification models to improve predictive accuracy for diverse driving scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine-learning models are trained only on previously encountered scenarios, then the model structure remains simple and training data requirements are limited, but predictive accuracy for rare or critical driving conditions deteriorates
Solution Approach 1:
The system performs preliminary actions by proactively searching for and collecting image data of rare driving scenarios before they are needed for prediction. Natural language queries are used to discover and gather training data for scenarios like slippery roads, emergency vehicles, and road work ahead of time, ensuring the model is prepared for these critical but infrequent conditions
Solution Approach 2:
The patent introduces an intermediary natural language processing system that acts as a mediator between the user's information needs and the image database. This intermediary enables efficient discovery and retrieval of relevant training data for rare scenarios without requiring manual searching or extensive data curation
2Reliability
If extensive training data for all possible scenarios is collected, then predictive accuracy for rare scenarios improves, but data processing time and computational resources increase
Solution Approach 1:
The system extracts only the essential and relevant image data needed for training by using natural language queries to identify specific rare scenarios. Instead of processing all available image data, the system selectively extracts training samples for critical scenarios like icy roads or emergency vehicles, reducing overall processing time while maintaining accuracy for important cases
Solution Approach 2:
The patent changes the parameter of data selection from comprehensive to targeted by introducing natural language query parameters. This allows the system to dynamically adjust which scenarios are prioritized for training based on their criticality, processing time-efficiently by focusing computational resources on the most important rare scenarios
Data Source
AI summary
A method for classifying environmental image frames for vehicles, using a universal iterative classification model, is presented. The method combines text-based querying, active machine-learning models, and user input to form an end-to-end automatic flow for sourcing video frames captured by sensors on a vehicle. An index of frames is used and continuously populated with data for new frames, with each frame scored on its likelihood of containing a representation of a driving scenario of interest. Each iteration of the classification model produces a classification result that is predicted to belong to the scenario of interest. Binary labels can be applied to the results. Subsequent iterative training of classification models can be performed using updated training sets containing previously labeled classification results, to improve precision and accuracy in classifying image data to a driving scenario.