Neural Network Safety Assessment for Automated Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural networks used in automated driving vehicles lack robustness and generalization, particularly in handling unseen environmental conditions such as rain or fog, and standard assessment metrics do not adequately quantify safety for perception tasks, which is critical for reliable vehicle operation.
Innovation Solution
A method is introduced to assess neural networks by measuring their accuracy and assigning a safety coefficient to each detected object based on its relevance, which considers distance and class, allowing for a weighted metric that evaluates network safety and importance, enabling the selection of the safest neural network for vehicle perception with low computational effort.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If standard performance metrics (mIoU, average precision) are used to assess neural networks, then the assessment is simple and computationally efficient, but the metrics do not adequately quantify safety for perception tasks
Solution Approach 1:
The patent transforms standard performance metrics into safety-relevant metrics by applying safety coefficients that weight different object classes and detection outcomes. This parameter transformation allows the same neural network output to be evaluated differently based on safety implications, converting ordinary accuracy measurements into safety-quantified metrics without requiring fundamentally new assessment infrastructure
2Reliability
If neural networks are trained to handle diverse real-world conditions and unseen corruptions, then robustness and generalization improve, but training stability decreases and computational resources increase
Solution Approach 1:
The patent applies data augmentation techniques during the training phase to pre-expose the neural network to various environmental conditions such as rain, fog, and other corruptions. By performing these augmentations in advance during training, the network learns to generalize to unseen conditions without requiring unstable retraining when encountering new scenarios in deployment
3Measurement precision
If safety coefficients are assigned to all detected objects to provide comprehensive safety assessment, then safety quantification accuracy improves, but computational effort increases
Solution Approach 1:
The patent applies different safety coefficients to different object classes and spatial regions based on their local importance to vehicle safety. Critical objects such as pedestrians, cyclists, and vehicles in proximity receive higher safety coefficients, while less critical objects receive lower coefficients. This localized differentiation provides accurate safety assessment for critical elements while reducing computational effort by not uniformly processing all detected objects with equal precision
Data Source
Figure 1
AI summary
The invention relates to a method for assessment of a neural network (NN) suitable for perception of an environment of a vehicle (10), in particular of an automated or autonomous driving vehicle, in order to find a neural network (NN), especially among a plurality of neural networks, which is safe for the use in the vehicle (10), the method comprising the following steps: - measuring an accuracy of the neural network (NN) in order to provide an accuracy metric (m) of the neural network (NN), - assigning a safety coefficient (γ) to the accuracy metric (m) of the neural network (NN) in order to provide a weighted metric (mr).