Autonomous Driving Reaction Models for Actual and Potential Risk Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current reaction models for autonomous driving do not adequately cover most dangerous driving scenarios, leading to improper evaluation of autonomous driving capability and compromised safety.
Innovation Solution
A reaction model building method that incorporates first and second risk points, where the first risk point is an actual threat and the second is a potential threat, allowing for the construction of models that accurately predict vehicle actions in various scenarios, including defensive and non-defensive emergency behaviors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a traditional reaction model is used for autonomous driving evaluation, then the model structure is simple and easy to implement, but it cannot cover most dangerous driving scenarios leading to improper evaluation
Solution Approach 1:
The reaction model is segmented into multiple specialized sub-models: a first reaction model for processing first risk points (actual threats) and a second reaction model for processing second risk points (potential threats). This segmentation allows each sub-model to be optimized for specific scenario types, improving overall coverage of dangerous driving scenarios while maintaining manageable complexity through modular architecture
Solution Approach 2:
The system dynamically selects which reaction model to apply based on the type of risk point detected. The evaluation process adapts its complexity by choosing the appropriate model (first or second) depending on whether the scenario involves actual threats or potential threats, allowing the system to scale computational resources according to scenario requirements
2Measurement precision
If the reaction model covers more dangerous scenarios with multiple risk points, then the evaluation accuracy improves, but the model complexity and calibration difficulty increase
Solution Approach 1:
The calibration process is segmented into distinct procedures for each reaction model. The first reaction model is calibrated using first calibration data from actual threat scenarios, while the second reaction model is calibrated using second calibration data from potential threat scenarios. This segmented calibration approach simplifies the overall process by allowing independent optimization of each model's parameters without requiring simultaneous calibration of the entire system
Solution Approach 2:
The system performs preliminary classification of risk points into two categories (actual threats and potential threats) before applying the appropriate reaction model. This preliminary action enables the system to pre-select the most suitable model for each scenario type, improving evaluation accuracy while avoiding the complexity of trying to calibrate a single universal model for all scenario types
3Ease of operation
If a single reaction model is used for all scenarios, then the model is easy to calibrate, but it cannot accurately predict vehicle actions in diverse dangerous scenarios
Solution Approach 1:
The system segments the prediction task into two distinct pathways: one for actual threats using the first reaction model, and another for potential threats using the second reaction model. Each model is calibrated independently using scenario-specific data, making the calibration process more manageable while significantly improving prediction accuracy for diverse dangerous scenarios compared to a single universal model
Solution Approach 2:
Different parameter sets are used for calibrating the first and second reaction models, optimized for their respective scenario types. The first model uses parameters derived from actual threat data, while the second model uses parameters from potential threat data. This parameter differentiation allows each model to be precisely tuned for its specific purpose, improving overall prediction accuracy while keeping the calibration process organized and manageable
Data Source
Figure 1
Figure 2~3A
Figure 3B
AI summary
A reaction model building method, a parameter calibration method for a reaction model, and a related apparatus are disclosed, and are applied to the field of intelligent driving technologies. The method includes: obtaining first information, and building a reaction model based on the first information. The first information includes information about a first risk point or a second risk point, the first risk point is an actual risk point that occurs on a road on which a first vehicle is located at a current moment, the second risk point is a potential risk point that occurs on the road on which the first vehicle is located at the current moment and that triggers a third risk point, and the third risk point is an actual risk point that occurs on the road on which the first vehicle is located in a first time period after the current moment. The reaction model indicates an action behavior of the first vehicle after the first risk point or the second risk point occurs. This method can cover most dangerous driving scenarios, thereby properly evaluating an autonomous driving capability, and improving autonomous driving safety.