Vehicle Safety Function Control Using Situation-Specific AI Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current safety-relevant functions in vehicles face challenges in producing predictable results due to the complexity of real-world input data, often resulting in infinite possibilities, known as the 'open world problem', which neural networks struggle to manage effectively.
Innovation Solution
A method utilizing multiple evaluation units trained to a predefined confidence level for specific driving situations, where input data is evaluated by a prioritized unit to execute safety-relevant functions, such as pedestrian detection or road condition assessment, ensuring reliable and precise decision-making.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If neural networks are used for detection and classification tasks in image data, then the ability to handle complex real-world input data is improved, but the predictability and reliability of safety-relevant functions deteriorates due to the open world problem and infinite possibilities of input data variations
Solution Approach 1:
The patent divides the evaluation of input data into multiple specialized neural networks, each trained for a specific driving situation (e.g., urban environment, highway, parking). This segmentation allows each network to focus on a limited set of scenarios, improving predictability within its domain while collectively covering the full range of driving situations. The system selects and activates only the relevant network based on the current driving situation, reducing the effective input space each network must handle.
2Reliability
If multiple specialized neural networks are deployed for different driving situations, then the reliability and predictability of safety-relevant functions is improved, but the device complexity and computational resources required deteriorates
Solution Approach 1:
The patent implements a dynamic selection mechanism that activates only the neural network relevant to the current driving situation. The system continuously monitors driving conditions and dynamically switches between pre-trained networks, ensuring that only the necessary computational resources are active at any given moment. This dynamic approach maintains high reliability by using the most appropriate specialized network while managing device complexity through selective activation rather than continuous operation of all networks.
3Device complexity
If a single comprehensive neural network is used to handle all possible driving situations, then the device complexity is reduced, but the measurement precision and reliability for specific driving situations deteriorates due to the vast input space
Solution Approach 1:
The patent applies local quality by training each neural network to specialize in specific local conditions or driving situations (e.g., urban traffic, rural roads, adverse weather). Each network develops optimized features and decision boundaries tailored to its specific domain, achieving high measurement precision for its targeted situations. This contrasts with a single comprehensive network that would need to generalize across all conditions, potentially sacrificing precision in any specific domain.
Data Source
AI summary
Technologies and techniques for executing a safety-relevant function of a vehicle in a current driving situation) of the vehicle depending on input data, which are evaluable by a control system with multiple evaluation units configured to evaluate the input data utilizing artificial intelligence, in which the evaluation units are trained by training data at least to a predefined confidence level in each case for a specific driving situation. An associated computer program product and a vehicle utilizing the technologies and techniques are further disclosed.