Traffic Light Detection Augmentation via GAN Superposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traffic light detection in autonomous driving applications is challenging due to the small size, distance, and low contrast of traffic lights in image data, requiring large amounts of labeled training data that are scarce.
Innovation Solution
Data augmentation methods are employed to generate augmented roadway scene images by superimposing traffic light images onto real images, using techniques like Generative Adversarial Networks (GANs) to determine likely locations and dimensions, increasing the realism and semantic consistency of training data for deep neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is used to generate training data, then the quantity of training data is improved, but the manufacturing precision (realism and quality) deteriorates
Solution Approach 1:
The patent uses GANs as an intermediary system between real images and augmented training data. The GAN consists of a generator that creates synthetic traffic light images and a discriminator that evaluates their realism. This intermediary process enables the generation of high-quality, realistic augmented images that maintain visual authenticity while increasing training data quantity.
Solution Approach 2:
The patent applies parameter changes by modifying various attributes of traffic lights in the augmented images, including position, size, color, orientation, and environmental conditions. These parameter variations are controlled through the GAN framework to generate diverse yet realistic training samples that improve model generalization without sacrificing image quality.
2Measurement precision
If deep neural networks are used to detect traffic lights, then the detection precision is improved, but the loss of information (due to overfitting) worsens when training data is scarce
Solution Approach 1:
The patent applies preliminary action by pre-training the deep neural network on a large dataset of augmented images generated by GANs before fine-tuning on real data. This preliminary exposure to diverse synthetic scenarios prepares the model to handle various traffic light conditions, reducing overfitting when trained on limited real labeled data and improving generalization performance.
3Measurement precision
If large amounts of real labeled training data are collected, then the detection precision is improved, but the ease of manufacture (data collection difficulty) deteriorates
Solution Approach 1:
The patent uses copying by generating synthetic copies of traffic light images through GANs instead of collecting additional real labeled images. The GAN learns from a small set of real images and generates numerous realistic copies with varied characteristics, providing abundant training data without the logistical challenges of real data collection, annotation, and storage.
Data Source
AI summary
Methods, systems, and non-transitory computer-readable media for generating augmented data to train a deep neural network to detect traffic lights in image data. The method includes receiving a plurality of real roadway scene images and selecting a subset of the plurality of real roadway scene images. The method also includes selecting an image from the subset and determining a distribution indicting how likely each location in the selected image can contain a traffic light. The method further includes selecting a location in the selected image by sampling the distribution and superimposing a traffic light image onto the selected image at the selected location to generate an augmented roadway scene image. The method also includes processing each image in the subset to generate a plurality of augmented roadway scene images. The method further includes training a deep neural network model using the pluralities of real and augmented roadway scene images.