Vehicle Image Data Generation for Autonomous Driving Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The performance of neural networks used for autonomous driving is limited by the quantity and quality of training data, leading to suboptimal performance in complex transportation scenarios.
Innovation Solution
An electronic device equipped with a camera, processor, and communication circuitry executes models to detect subjects from images, generate enhanced training data by processing feature information, and selectively transmit data to a server for neural network training, using generative adversarial networks to improve data quality and quantity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If more training data is collected to improve neural network performance, then the quantity of training data increases, but the complexity of data collection and processing increases
Solution Approach 1:
The patent uses generative adversarial networks to create synthetic training data that copies the essential characteristics of real driving scenarios. The GAN generates artificial images of road scenes, pedestrians, vehicles, and environmental conditions that can be used to train autonomous driving models without requiring physical collection of equivalent real-world data, thus increasing training data quantity while avoiding the complexity of extensive field data collection.
Solution Approach 2:
The patent introduces an intermediary processing system that includes image processing modules, feature extraction modules, and GAN-based generation modules. This intermediary layer transforms limited real training data into expanded synthetic datasets, mediating between the constraint of limited real data and the need for large-scale training data, thereby reducing the direct complexity of data collection while achieving data quantity goals.
2Manufacturing precision
If the quality of training data is improved to enhance neural network performance, then the quality of training data increases, but the time and resources required for data selection and processing increase
Solution Approach 1:
The patent applies preliminary filtering and selection processes to identify high-quality real training data before feeding it into the GAN system. By pre-processing and selecting only the most valuable real-world driving scenarios, the system ensures that the synthetic data generated builds upon a solid foundation of quality, thereby achieving high overall training data quality while minimizing the time spent on extensive manual data curation.
Solution Approach 2:
The patent transforms training data quality through parameter changes in the GAN generation process, adjusting factors such as image resolution, environmental conditions, object positions, and physical parameters to create diverse yet realistic training scenarios. This automated parameter variation allows the system to generate high-quality diverse training data without manual intervention, reducing the time required for data quality enhancement.
3Quantity of substance
If generative adversarial networks are used to generate synthetic training data, then the quantity and quality of training data improve, but the computational complexity and processing time increase
Solution Approach 1:
The patent segments the GAN training process into multiple stages: pre-training phase with real data, fine-tuning phase with mixed real and synthetic data, and deployment phase. This segmentation allows the computationally intensive GAN training to be performed separately from the autonomous driving model training, distributing computational complexity across different time periods and hardware resources, thereby making the overall system more manageable.
Solution Approach 2:
The patent employs partial action by generating only the specific types and quantities of synthetic data needed for particular training scenarios rather than generating all possible training data. The system selectively generates synthetic data for underrepresented scenarios (e.g., rare weather conditions, edge cases) while relying on real data for common scenarios, thus reducing unnecessary computational complexity while still achieving data quantity and diversity goals.
Data Source
AI summary
An electronic device in a vehicle may comprise communication circuitry, a camera, a memory storing instructions, and a processor. The instructions may, when executed by the processor, cause the electronic device to execute a first model to detect one or more subjects from a first image obtained from the camera and obtain, by executing a second model using feature information obtained from the first model, a second image based on the feature information.