Autonomous Driving Model Scenario Adaptation via GAN Domain Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving methods using the end-to-end approach have limited application scenarios and low robustness, as models trained in specific conditions cannot perform effectively in different scenarios, such as daytime vs. nighttime or sunny vs. rainy conditions, or when pedestrians and vehicles are present vs. absent.
Innovation Solution
The method involves using a pre-trained autonomous driving model that can indicate relationships between images and driving data across multiple scenarios, including generating sample images using domain conversion models, such as those trained with generative adversarial networks (GAN), to adapt driving data for various conditions like daytime to nighttime or sunny to rainy, and adding or removing vehicles and pedestrians in images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a model is trained using end-to-end method with data collected in a certain scenario, then the model can be obtained for that specific scenario, but the model cannot be used in other scenarios resulting in limited application scenario and low robustness
Solution Approach 1:
The patent applies preliminary action by pre-training an autonomous driving model using data from multiple scenarios before actual deployment. The model is preliminarily exposed to various conditions (daytime, nighttime, sunny, rainy, different obstacle configurations) during the training phase, so that when the vehicle encounters any of these scenarios during operation, the model already has pre-acquired knowledge to handle them effectively, eliminating the need for scenario-specific model switching
Solution Approach 2:
The patent implements universality by creating a single autonomous driving model that can handle multiple scenarios universally. Instead of having separate models for different scenarios, the trained model functions across all scenarios (different times, weather conditions, obstacle configurations), making the system multi-functional and adaptable to various driving conditions without requiring scenario-specific adaptations
2Reliability
If data is collected in specific conditions for model training, then the model performs well in those conditions, but it cannot produce appropriate output in different conditions such as daytime vs nighttime or sunny vs rainy
Solution Approach 1:
The system performs preliminary action by collecting training data from diverse conditions (daytime, nighttime, sunny, rainy, various obstacle configurations) and training the model in advance. This preliminary exposure to multiple scenarios ensures that when the vehicle operates in any of these conditions, the model has already learned the appropriate responses, maintaining reliable performance across all scenarios without requiring real-time adaptation
Solution Approach 2:
The patent applies parameter changes by varying multiple parameters during data collection and model training, including time of day, weather conditions, and obstacle configurations. By training the model with data spanning different parameter values, the system learns to adapt its output based on the current parameter state, enabling reliable performance across varying conditions such as daytime versus nighttime or sunny versus rainy weather
3Productivity
If the autonomous driving model is trained with limited scenario data, then the training process is simpler and faster, but the model lacks robustness when encountering unseen scenarios
Solution Approach 1:
The system applies preliminary action by comprehensively preparing training data from multiple scenarios before model training. Instead of training incrementally or requiring real-time data collection, all necessary training data from various scenarios is prepared in advance, allowing the model to learn robust patterns across different conditions during a single training process, thereby achieving both training efficiency and model robustness
Solution Approach 2:
The patent applies segmentation by dividing the training process into distinct scenario categories (daytime, nighttime, sunny, rainy, different obstacle configurations). Each scenario type is represented in the training data, allowing the model to learn specific patterns for each segment while integrating them into a unified model. This segmented approach to training ensures comprehensive coverage without requiring separate training processes for each scenario
Data Source
AI summary
The present disclosure provides an autonomous driving method and an apparatus. The method includes: receiving a currently collected image transmitted by a unmanned vehicle, where the currently collected image is an image collected in a target scenario; acquiring current driving data according to the currently collected image and a pre-trained autonomous driving model, where the autonomous driving model is used to indicate a relationship between an image and driving data in at least two scenarios, and the at least two scenarios include the target scenario; and sending the current driving data to the unmanned vehicle. Robustness of the unmanned driving method is improved.