In recent years, with the spread of deep learning, research for the practical use of autonomous driving has been actively promoted. In order to support safe driving, it is important to understand extraordinary conditions such as road abnormalities during disasters and traffic congestion caused by events. Therefore, in this research, we recognize the abnormal state on the road using deep learning.
Because such unusual events rarely occur, the amount of training data for deep learning is usually insufficient. Therefore, we aim to solve the problem by generating data of extraordinary events using deep learning.
Generate flooded image
Multi-domain image transformation using semantic information
We propose a method for image transformation using semantic information of images and multiple road states when generating extraordinary scenes using GANs. At the time of image transformation, labels are added to input information to process multiple domains. Generator learns image transformations so that Discriminator classifies them as the domain specified in the input. In addition, by adding a constraint so that semantic information does not change, unexpected transformation is suppressed.
Fig. 1: Training Generator
Fig. 2: Training Discriminator
Fig. 3: Image transformation example (From left: original image, generated image, original image, generated image)
Flooded scene detection
In order to show the effect of data augmentation, extraordinary scenes were detected by adding generated images to extraordinary scenes. 100images has no data extension. Others are the results of similar data expansion using various methods.
Result of flooded detection
Publications
|
