Driving Scene Token Selection for Efficient Road Element Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Vehicles with autonomous driving capabilities and/or driver assistance capabilities face challenges in processing real-time information about road elements efficiently and effectively.
Innovation Solution
Utilizing transformer neural networks for processing road element information, employing multiple tokens per element, including classification, location, and behavioral tokens, and using attention mechanisms to identify relevant tokens, thereby simplifying processing and reducing resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional methods are used to process road element information in real-time, then comprehensive information can be captured, but processing efficiency and resource consumption are insufficient
Solution Approach 1:
The patent segments road element information processing into multiple token types (classification tokens, location tokens, behavioral tokens). Each token represents a specific aspect of the road element, allowing the system to process only relevant information segments rather than treating all data uniformly, thereby improving processing efficiency while reducing unnecessary resource consumption.
Solution Approach 2:
The patent extracts and processes only the most relevant tokens for each road element based on the driving scenario. By identifying and focusing on critical information (such as behavioral tokens for dynamic objects or location tokens for spatial relationships), the system eliminates processing of redundant data, achieving both high processing efficiency and reduced resource consumption.
2Reliability
If all tokens representing road elements are processed in detail, then comprehensive understanding is achieved, but processing complexity and resource allocation increase
Solution Approach 1:
The patent dynamically adjusts which tokens are processed based on the driving scenario and context. The system can adaptively select and prioritize tokens relevant to the current situation (e.g., focusing on behavioral tokens during high-risk scenarios or location tokens during navigation), maintaining high processing accuracy while reducing overall processing complexity through context-aware selective processing.
Solution Approach 2:
Different token types receive different processing priorities and computational resources based on their relevance to the driving task. Classification tokens may receive simpler processing while behavioral tokens requiring prediction receive more sophisticated analysis. This localized quality approach ensures accurate processing of critical information while reducing complexity for less critical data.
3Measurement precision
If multiple tokens per road element are used to represent different attributes, then granular information is obtained, but data volume and processing load increase
Solution Approach 1:
The patent applies partial processing by selecting and processing only the necessary subset of tokens for each driving scenario. Rather than processing all tokens uniformly, the system performs partial action on the most relevant tokens (e.g., processing behavioral tokens for dynamic objects while using simpler processing for static elements), achieving sufficient measurement precision without the full data volume burden.
Solution Approach 2:
The token system is designed to be universal and multi-functional, where the same token structure can represent different aspects of road elements depending on the context. This allows the system to achieve granular representation when needed while consolidating data representation when full granularity is not required, effectively managing data volume while maintaining measurement precision on demand.
Data Source
AI summary
A method of providing selective learning by prediction for driving, the method includes (a) obtaining, by a machine learning process using an artificial neural network trained across road elements, a first set of tokens with respect to an element captured in a sensed information unit in an environment of a vehicle, the first set of tokens representing respective attributes characterizing the first element; (b) obtaining, by the machine learning process, a second set of tokens generated respect to the vehicle and representing respective attributes characterizing the vehicle; (c) obtaining, by the machine learning process, a scenario indication that is indicative of a scenario faced by the vehicle in the environment; and (d) processing, by the machine learning process, the first set of tokens in correspondence with the second set of tokens and with respect to the scenario.