Autonomous Vehicle Image Augmentation for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training and testing of autonomous transportation vehicles face challenges in adequately preparing neural networks for real-world scenarios, particularly in handling undesirable situations like veering off-lane, due to limitations in available image data and practical considerations such as data storage and the rarity of dangerous scenarios during human-driven testing.
Innovation Solution
The system employs image augmentation techniques to simulate various driving scenarios by modifying captured images, allowing the neural network to train on a broader range of conditions, including remedial courses of travel, thereby enhancing autonomous vehicle operations without the need for extensive real-world data collection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real-world image data is collected during human-driven testing, then the training data reflects actual driving scenarios, but the quantity and diversity of training data is insufficient due to the rarity of dangerous scenarios
Solution Approach 1:
The patent creates synthetic copies of real driving images by generating virtual images that simulate dangerous driving scenarios. These virtual images are synthesized by modifying existing real-world images to introduce hazardous conditions such as pedestrians, cyclists, and vehicles in dangerous positions, thereby multiplying the training data without requiring actual collection of rare dangerous scenarios.
Solution Approach 2:
The patent modifies image parameters to generate diverse training scenarios. By changing parameters such as object positions, lighting conditions, weather conditions, and scene configurations in virtual images, the system creates a wide variety of training data from limited real-world samples, addressing both quantity and diversity requirements.
2Adaptability or versatility
If extensive real-world data collection is conducted to cover all driving scenarios, then the neural network can be trained on diverse conditions, but the data storage requirements and testing time become excessive
Solution Approach 1:
The patent performs preliminary synthesis of virtual training images before actual neural network training. By pre-generating a comprehensive set of virtual driving scenarios that cover diverse and dangerous conditions, the system prepares all necessary training data in advance, eliminating the need for time-consuming real-world data collection during the testing phase.
Solution Approach 2:
The virtual image generation system serves multiple functions: it creates training data for diverse scenarios, simulates rare dangerous conditions, and provides a scalable solution that can generate unlimited variations. This single virtual environment replaces the need for extensive real-world data collection across multiple test sites and conditions.
3Productivity
If the neural network is trained only on normal driving conditions, then the training process is simpler and faster, but the system fails to handle undesirable situations like veering off-lane
Solution Approach 1:
The patent applies preliminary anti-action by pre-training the neural network on virtual images that explicitly depict dangerous and undesirable driving scenarios. By exposing the network to synthesized hazardous conditions before deployment, the system prepares the neural network to recognize and respond to dangerous situations, countering the limitation of training only on normal conditions.
Data Source
AI summary
Devices, systems, and methods related to autonomous transportation vehicle operation may include image augmentation arrangements for training and/or evaluating autonomous operations. Such augmentations may include artificial impressions of driving conditions which can prompt recovery operations.