Autonomous Vehicle Mode Switching via On-Device CNN-RPN Confidence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving technologies face challenges in switching from autonomous to manual mode effectively, particularly in unpredictable conditions such as sudden darkness or broken street lights, due to the limitations of passively updated geographic zones databases, which fail to accurately determine safe driving scenarios.
Innovation Solution
A learning method and device using a Convolutional Neural Network (CNN), Region Proposal Network (RPN), loss layer, and confidence layer to generate credibility parameters for object detection, enabling on-device standalone predictions for mode switching in autonomous vehicles, ensuring safety by integrating CNN and RPN confidence scores for accurate object detection and mode switching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a geographic zones database is used to determine safe driving sections, then the autonomous vehicle can operate in predefined safe zones, but it cannot adapt to unexpected conditions such as sudden darkness or broken street lights
Solution Approach 1:
The system performs preliminary object detection and environmental assessment before autonomous driving begins, using CNN and RPN to pre-evaluate the safety of the current environment. This allows the vehicle to proactively identify potential hazards in advance and prepare for mode switching before adverse conditions actually occur, rather than reacting passively after the fact.
Solution Approach 2:
The system continuously monitors environmental conditions and object detection results during autonomous driving, using the confidence scores from CNN and RPN as feedback signals. When the feedback indicates deteriorating conditions (e.g., confidence scores drop below thresholds due to darkness or obstacles), the system automatically triggers mode switching from autonomous to manual driving, creating a closed-loop safety mechanism.
2Ease of manufacture
If passive database updates are used for geographic zones, then the system is simple to maintain, but it cannot reflect real-time changes in environmental conditions
Solution Approach 1:
The autonomous vehicle performs self-assessment of its operating environment using onboard CNN and RPN systems. Instead of relying entirely on externally maintained geographic zone databases, the vehicle independently evaluates current conditions through object detection, confidence scoring, and real-time environmental analysis, enabling it to make autonomous safety decisions without requiring frequent database updates.
Solution Approach 2:
The patent replaces the mechanical/database-dependent approach of geographic zone lookup with an intelligent system using deep learning models (CNN and RPN). Instead of querying a static database, the system uses neural networks to process visual input and dynamically assess safety, substituting computational intelligence for traditional database querying mechanisms.
3Reliability
If manual driving mode is required frequently due to database limitations, then safety may be improved, but driving efficiency and productivity decrease
Solution Approach 1:
The system applies object detection and confidence scoring selectively to specific regions of interest identified by RPN, rather than analyzing the entire visual field uniformly. This partial action approach focuses computational resources on critical areas, maintaining high safety monitoring while reducing overall processing time and enabling faster decision-making for mode switching.
Solution Approach 2:
The system dynamically adjusts confidence score thresholds and detection parameters based on environmental conditions and driving context. By changing these parameters adaptively, the system can maintain high safety standards while avoiding excessive manual mode switching in low-risk situations, thereby preserving driving efficiency while ensuring safety.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A learning method for generating parameters capable of representing a degree of credibility of an object detection during a process of the object detection is provided. And the method includes steps of: (a) a learning device instructing a convolutional layer to generate a convolutional feature map by applying a convolutional operation to a training image; (b) the learning device instructing an anchor layer to generate an RPN confidence map including RPN confidence scores; (c) the learning device instructing an FC layer to generate CNN confidence scores, to thereby generate a CNN confidence map; and (d) the learning device instructing a loss layer to learn parameters in the CNN and the RPN by performing backpropagation using an RPN loss and a CNN loss, generated by referring to the RPN confidence map, the CNN confidence map, an estimated object detection result and a GT object detection result.