Automotive Vision OOD Detection for Reliable Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated vehicles rely on sensor data for safe operation, but existing systems lack effective methods to ensure the reliability of object classification and instance detection, which is crucial for timely and safe control of driving automation features.
Innovation Solution
A method using an automotive vision system with an input encoder, internal autoencoder, object class decoder, and instance decoder to determine reliability through an out-of-distribution (OOD) indicator and uncertainty indicator, ensuring reliable object detection and classification by assessing the validity of sensor data in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If automotive sensor data is used for object detection and classification in automated vehicles, then the vehicle can perform driving automation features, but the reliability of sensor data and object classification cannot be ensured
Solution Approach 1:
The system implements feedback by continuously monitoring the reliability of sensor data and object classification results through OOD indicators and uncertainty measurements. When reliability thresholds are violated, the system feeds back control signals to switch to minimal risk conditions, creating a closed-loop reliability assurance mechanism that maintains safe operation while enabling automation.
Solution Approach 2:
The patent introduces intermediary components including an internal autoencoder that generates OOD indicators and a reliability determination module that assesses classification confidence. These intermediaries act as mediators between the raw sensor data and the driving automation system, filtering and evaluating data quality before it influences automated driving decisions.
2Reliability
If the vehicle ensures reliability of sensor data quickly, then timely safety measures can be taken, but the complexity of the vision system increases
Solution Approach 1:
The system merges the reliability determination functionality directly into the existing automotive vision system architecture. The internal autoencoder, object class decoder, and instance decoder are integrated components that simultaneously perform object recognition and reliability assessment, eliminating the need for separate verification systems and reducing overall complexity.
Solution Approach 2:
The automotive vision system is designed with multi-functionality, where the same neural network components (input encoder, internal autoencoder, decoders) serve dual purposes: both performing object detection and classification while simultaneously generating reliability metrics through OOD indicators and uncertainty measurements. This universal approach avoids duplicating system components.
3Reliability
If the vehicle ensures reliability of object classification, then safe control is improved, but the processing time and computational load increase
Solution Approach 1:
The internal autoencoder performs preliminary action by generating OOD indicators during the feature extraction phase, before final classification decisions are made. This preliminary reliability assessment allows the system to quickly identify unreliable inputs without requiring additional processing time after classification, enabling timely safety responses.
Solution Approach 2:
The reliability determination process operates continuously alongside object detection and classification without interrupting the main processing flow. The neural network components continuously generate both classification results and reliability metrics in parallel, ensuring that safety assessments are always available without adding sequential processing delays.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present disclosure relates to determining a reliability of object instance detection and classification performed by an automotive vision system on sets of automotive sensor data deployed in an at least partially automated vehicle. Based on a latent space vector of an internal autoencoder of the automotive vision system and a corresponding autoencoder latent space probability density function, an out of distribution (OOD) indicator is determined for each set of automotive sensor data. If a comparison of the OOD indicator of a sensor data reliability number of subsequent sets of automotive sensor data indicates a violation of an OOD threshold, a driving automation system feature of the vehicle is caused to achieve a minimal risk condition of the vehicle.