CNN Training With Synthetic Driving Data for Corner Cases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning techniques for autonomous driving struggle to generalize across various driving scenarios and require extensive training data, often leading to biased performance in real-world environments, especially in complex situations like intersections, and lack effective methods for replicating human-like driving behaviors.
Innovation Solution
A hybrid training approach using real-world and synthetic driving data, combined with a genetic algorithm, to train a convolutional neural network, where synthetic data is generated using a generative process modeling human-like driving behaviors, and a multi-objective neuroevolutionary method is employed to optimize the network's performance based on a Pareto front of fitness vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If end-to-end deep learning is used to map raw input data to control signals, then the system can learn complex driving strategies, but it requires a very high capacity model and a large amount of training data to handle corner cases
Solution Approach 1:
The patent segments the training data into real-world driving data and synthetic driving data. The synthetic data is generated to specifically cover corner cases and rare scenarios that are difficult to capture in real-world data, allowing the model to learn complex strategies without requiring excessively large amounts of real training data.
Solution Approach 2:
The patent uses synthetic data generation to create copies of driving scenarios that replicate real-world conditions while enabling controlled generation of rare corner cases. This allows the model to be trained on diverse scenarios without requiring proportionally large amounts of real-world data collection.
2Adaptability or versatility
If deep reinforcement learning is used to maximize cumulative reward through environment exploration, then the system can learn driving behaviors, but it tends to learn a biased version of the driving environment and performs poorly when ported to real-world vehicles
Solution Approach 1:
The patent introduces synthetic driving data as an intermediary between simulated reinforcement learning and real-world deployment. The synthetic data is generated to match real-world distributions more closely, acting as a bridge that reduces the domain gap and improves transferability to real vehicles.
Solution Approach 2:
The patent changes the parameters of the training environment by using synthetic data generation with realistic distributions rather than purely simulated environments. This parameter change in the training data characteristics helps the model generalize better to real-world conditions.
3Adaptability or versatility
If a high capacity model is used with end-to-end learning to handle corner cases, then the system can learn multi-path situations, but the device complexity increases
Solution Approach 1:
The patent segments the learning task by using synthetic data to pre-train the model on corner cases and rare scenarios before fine-tuning with real-world data. This segmentation of training phases allows the model to develop specialized knowledge for corner cases without requiring excessive model capacity throughout the entire training process.
Data Source
AI summary
This disclosure is related to a method, a computer program code, and an apparatus for training a convolutional neural network for an autonomous driving system. The disclosure is further related to a convolutional neural network, to an autonomous driving system comprising a neural network, and to an autonomous or semi-autonomous vehicle comprising such an autonomous driving system. For training the convolutional neural network, in a first step real-world driving data are selected as training data. Furthermore, synthetic driving data are generated as training data. The convolutional neural network is then trained on the selected real-world driving data and the generated synthetic driving data using a genetic algorithm.