ADAS Training Data Generation Using GAN-Simulated Road Conditions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing training devices face inefficiencies in collecting comprehensive road condition data due to climate and regional limitations, leading to ADAS models that struggle to adapt to various road conditions, affecting driving safety.
Innovation Solution
A training set generation method using a Generative Adversarial Network (GAN) to simulate driving scenarios, combining an initial generator and discriminator to generate data that mimics real-world conditions, reducing the need for extensive real data collection and enhancing model training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training devices collect real driving data from actual road conditions, then the training data reflects authentic driving scenarios, but the collection efficiency is low and comprehensive coverage is limited due to climate and regional constraints
Solution Approach 1:
The patent uses a generative adversarial network to create synthetic driving data that copies the statistical properties and visual characteristics of real driving data. The generator network produces images that mimic real road scenes, while the discriminator network ensures they are indistinguishable from authentic captured images, thus resolving the contradiction between data authenticity and collection efficiency
Solution Approach 2:
The patent replaces the mechanical data collection process (physical sensors capturing real-world scenes) with a computational generation process. Instead of using cameras and sensors to physically capture driving data under various conditions, the system uses neural networks to synthetically generate driving scenarios, eliminating the limitations of physical data collection
2Adaptability or versatility
If training devices collect comprehensive driving data across various road conditions, then the ADAS model can handle diverse scenarios, but the time and resources required for data collection increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-training the generative model on a diverse set of driving conditions. Once trained, the generator can rapidly produce synthetic data for any desired road condition without requiring actual field collection. This preliminary training phase enables fast generation of comprehensive training datasets for various scenarios including different weather, lighting, and road types
Solution Approach 2:
The patent utilizes parameter changes in the generative model to efficiently generate diverse driving conditions. By adjusting input parameters such as weather conditions, time of day, road type, and traffic density, the system can generate comprehensive training data for various road conditions without physical re-collection, thus improving adaptability while reducing time loss
Data Source
AI summary
A training set generation method comprises generating simulated driving data based on an advanced driving assistance system (ADAS) simulator, inputting the simulated driving data into an objective generator to obtain generated data, wherein the objective generator is a generator trained by a generated adversarial network, and determining a model training set of an ADAS model based on the generated data. An electronic device is also disclosed.