Vehicle Obstacle Recognition Using Uncertainty-Based Travel Decisions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous driving systems face difficulties in recognizing obstacles during adverse conditions, leading to potential accidents due to incorrect control logic operation, especially when weather or edge cases complicate image recognition.
Innovation Solution
A device and method utilizing a convolutional neural network to extract objects, estimate center points and uncertain regions, and determine travel conditions based on uncertainty values, preventing abnormal control by generating warnings for caution or impossibility through a camera, detection device, center calculation device, uncertainty determination device, and condition determination device.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning technology is used for object recognition in autonomous driving, then recognition capability is improved, but reliability deteriorates under bad travel conditions due to uncertainty in extraction results
Solution Approach 1:
The system calculates uncertainty values for each extracted object and feeds this uncertainty information back into the control decision-making process. The condition determination device uses the uncertainty value to assess whether to trust the deep learning extraction result, creating a feedback loop that improves reliability without sacrificing recognition capability.
Solution Approach 2:
An uncertainty calculation module is introduced as an intermediary between the deep learning object recognizer and the control system. This intermediary computes uncertainty values that mediate the trust relationship, allowing the system to appropriately weight deep learning results based on their reliability rather than blindly following or completely rejecting them.
2Measurement precision
If separate training data for bad condition determination is collected, then determination accuracy is improved, but device complexity increases
Solution Approach 1:
The deep learning object recognizer serves itself by providing uncertainty information about its own extraction results. Instead of requiring separate training data and systems to determine bad conditions, the existing object recognition model generates uncertainty values that directly indicate its own confidence level, eliminating the need for additional data collection infrastructure.
Solution Approach 2:
The deep learning object recognizer is designed to perform multiple functions: not only does it extract objects from images, but it also generates uncertainty values that serve as bad condition indicators. This multi-functionality eliminates the need for separate bad condition detection systems and their associated training data requirements.
Data Source
AI summary
A device for recognizing an obstacle of a vehicle includes a camera for acquiring an image, a detection device for extracting an object by applying a convolutional neural network to the image, a center calculation device for estimating a center point and a region outside the center point of the extracted object, an uncertainty determination device for determining whether the estimated center point and region outside the center point are uncertain, and a condition determination device for determining whether travel is possible based on the determined uncertainty.