Vehicle Neural Network Training for Rare Scenario Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks for autonomous vehicles face challenges in handling rare scenarios due to imbalanced training data, leading to biased models that struggle with detecting critical 'corner cases' like pedestrians on freeways, as conventional data augmentation techniques fail to recreate realistic scenarios effectively.
Innovation Solution
The use of Generative Adversarial Networks (GANs) and computer graphics simulation to generate realistic image and video data, focusing on both local and global features, allows for the creation of new scenarios that can augment the training data, addressing the long tail problem by increasing the representation of rare events and improving model robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more real-world driving data is collected to train the neural network model, then the model can handle more common scenarios, but the training data becomes increasingly imbalanced with rare corner cases still underrepresented
Solution Approach 1:
The patent uses Generative Adversarial Networks (GANs) to create synthetic copies of corner case scenarios. The GAN generates realistic images and videos of rare events (e.g., pedestrians on freeways, animals on roads) that can be added to the training dataset, effectively copying the essential characteristics of rare scenarios without requiring physical collection of such events.
Solution Approach 2:
The system performs preliminary identification of corner cases by analyzing the training dataset to detect underrepresented scenarios. Before full model training, the system identifies which rare scenarios need more representation and generates targeted synthetic data for those specific cases, allowing proactive addressing of data imbalances.
2Adaptability or versatility
If conventional data augmentation techniques are used to increase training data diversity, then more scenarios can be covered, but the generated data lacks realism and fails to effectively improve corner case detection
Solution Approach 1:
Instead of using conventional augmentation techniques that simply transform existing images, the patent employs GANs to generate entirely new synthetic images and videos that copy the statistical properties and visual characteristics of real corner case data. This produces realistic-looking training samples that conventional methods cannot achieve.
Solution Approach 2:
The GAN architecture transforms the parameters and characteristics of input data through adversarial training, converting latent space representations into realistic images. The generator and discriminator networks work together to adjust parameters such as lighting, texture, and object placement to produce highly realistic synthetic corner case scenarios.
3Reliability
If the neural network model is trained exclusively on real collected data, then the model learns from authentic scenarios, but the training process is time-consuming and computationally expensive
Solution Approach 1:
The patent creates synthetic copies of real corner case scenarios using GANs, which can be generated much faster than collecting equivalent real-world data. This synthetic data serves as a substitute for time-consuming physical data collection while maintaining the essential characteristics needed for effective model training.
Solution Approach 2:
The system performs preliminary analysis of the training dataset to identify corner cases that need augmentation, then generates synthetic data specifically for those identified scenarios. This targeted approach avoids the need to collect and process large amounts of additional real-world data, significantly reducing training time while focusing resources on the most critical underrepresented cases.
Data Source
AI summary
A computing system comprises a data storage and at least one processor communicatively coupled to the data storage. The at least one processor is configured to execute program instructions to cause the system to perform the following steps. A deep neural network (“DNN”) model is trained using training data. Next, additional scenes are determined based on the DNN model and the training data. The determined scenes are generated, and then used to augment the training dataset. The DNN model is then retrained using the augmented training dataset and stored in a data storage for deployment.