Autonomous Vehicle Mode Selection via Visual Outlier Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous drive systems for vehicles struggle to reliably classify scenarios outside their intended operation regions, making it difficult to select the appropriate driving mode for semi-autonomous or autonomous vehicles.
Innovation Solution
A method that uses a machine learning algorithm to reconstruct visual representations of a vehicle's environment and calculates a difference metric based on deviations between the original and reconstructed images, allowing for classification of scenarios as either known or unknown, thereby selecting the appropriate driving mode.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the system specifies each operation region where autonomous operation is not allowed, then the system can recognize known scenarios reliably, but it becomes impossible to account for all unknown scenarios due to environmental complexity
Solution Approach 1:
Instead of trying to specify all regions where autonomous operation is not allowed (negative definition), the patent inverts the approach by specifying only the training data regions and allowing the system to identify any scenario outside these regions as unknown. This inversion transforms an impossible task of exhaustive negative specification into a feasible task of positive training followed by outlier detection.
Solution Approach 2:
The patent uses a lightweight machine learning model trained on representative training data that can be quickly updated or retrained when new scenarios are encountered. Rather than maintaining a complex, exhaustive database of all possible operation regions, the system uses a simpler, more adaptable model that can be refreshed with new data as needed.
2Adaptability or versatility
If the system classifies scenarios as known or unknown, then appropriate driving mode can be selected, but the system cannot recognize unknown scenarios without extensive pre-programming of all possible operation regions
Solution Approach 1:
The patent performs preliminary training of the machine learning model using representative training data that captures the normal operation regions. This preliminary action prepares the system in advance to quickly classify new scenarios by comparing them against the trained model, eliminating the need for time-consuming analysis of all possible operation regions during actual operation.
Solution Approach 2:
The patent replaces the mechanical approach of explicitly programming all operation regions and boundaries with a data-driven machine learning approach. The system substitutes complex rule-based classification with a trained neural network or similar model that can generalize from training data to quickly classify both known and unknown scenarios.
3Measurement precision
If the system uses traditional scenario recognition methods, then it can operate within defined regions, but it cannot reliably identify scenarios outside the intended operation region
Solution Approach 1:
The patent extracts only the essential features from training data that are sufficient to define normal operation regions, rather than attempting to capture and process all possible scenario details. By taking out only the critical patterns from training data, the system achieves precise classification with reduced computational complexity.
Solution Approach 2:
The patent changes the parameters used for scenario recognition from explicit operation region boundaries to learned features from training data. By transforming the recognition parameters from predefined geographic or situational boundaries to data-driven feature representations, the system achieves higher precision in identifying both known and unknown scenarios.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The disclosed subject matter generally relates to a method for selecting an operation mode among a plurality of operation modes for an at least partly self-driving vehicle. The proposed method provides for selecting a driving mode based on, in contrast to prior art methods, concluding that a present scenario captured in a visual representation is part of an unknown scenario, i.e. that is not recognized. If the captured visual representation is determined to be part of a known scenario, so-called inlier data representing a normal region of input data that can be recognized, the driving mode may be selected accordingly. However, if the visual representation is determined to be part of the unknown scenario, so-called outlier data that is at least partly un-recognized, another driving mode may be selected accordingly. The proposed method provides the advantage that the unknown scenarios do not have to be recognized as such, it is sufficient to conclude that they do not belong to the known region of scenarios or data.