Cost-Based Decision Rule for Autonomous Vehicle Safety
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving systems face challenges in ensuring safety during classification tasks, as existing methods are not adequately equipped to differentiate between critical errors, such as mistaking a pedestrian for the road, and less severe errors, and are computationally intensive, making them impractical for automotive control units.
Innovation Solution
A cost-based decision rule using a cost matrix is applied to the output of an artificial neural network to prioritize safety, allowing for a safety-focused classification that adjusts class assignments based on the potential consequences of errors, thereby enhancing safety in autonomous driving.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If cost-sensitive training is applied to penalize certain perceptual errors more severely, then safety-related aspects are taken into account, but convergence is difficult to guarantee and complete retraining is required for adjustments
Solution Approach 1:
The patent applies dynamics by making the cost matrix adjustable and reconfigurable without requiring complete retraining of the neural network. The cost matrix can be dynamically modified to reflect changing safety priorities or legal frameworks, allowing the system to adapt to new requirements while maintaining the trained model parameters.
Solution Approach 2:
The patent segments the safety adjustment mechanism from the core neural network training. By separating the cost matrix (which encodes safety priorities) from the neural network weights, the system allows independent adjustment of safety parameters without affecting or requiring retraining of the underlying classification model.
2Reliability
If uncertainty assessment methods such as Monte Carlo dropout or posterior entropy calculation are used, then indication of incorrect classification is provided, but computational intensity is very high for automotive control units
Solution Approach 1:
The patent uses a computationally inexpensive cost matrix that can be quickly applied to neural network outputs without requiring intensive calculations. Instead of using heavy methods like Monte Carlo dropout or posterior entropy, the system applies a simple matrix multiplication that is computationally lightweight and suitable for real-time automotive applications.
Solution Approach 2:
The patent extracts the uncertainty and safety assessment function from computationally intensive methods and implements it through a simplified cost matrix approach. This extraction allows the system to maintain reliability assessment capabilities while dramatically reducing computational requirements for automotive control units.
3Productivity
If traditional Argmax function is used to interpret the posterior, then classification result is obtained quickly, but safety-related aspects are not adequately considered as all errors are treated equally
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different classification errors through the cost matrix. Instead of treating all errors equally as in traditional Argmax, the cost matrix assigns different weights to different types of errors based on their safety implications, allowing critical errors to be penalized more heavily while maintaining efficient processing.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
To improve safety in autonomous or semi-autonomous driving, the invention provides a method for automatically executing a control function of an autonomously or semi-autonomously controlled vehicle (100), which comprises the following steps: - Determining sensor data using at least one sensor (210) of a vehicle (100), - Processing the sensor data using an artificial neural network (500), wherein the sensor data or data derived from the sensor data are supplied to the artificial neural network (500) as input variables and the artificial neural network (500) maps the input variables to output variables depending on a parameterizable processing chain (410, 420, 430), - Classifying the determined sensor data depending on the output variables determined by the artificial neural network and depending on a predefined safety-oriented decision rule.wherein the safety-oriented decision rule is designed as a cost-based decision rule, and - execution of the control function depending on the classification. Furthermore, the invention provides a control system designed for executing the method with an evaluation unit, as well as a vehicle with such a control system.