Előadás címe: Synthetic Data Generation for Blocked Video Sensor
Időpont: 2020.06.30., 09:20 – 09:40
Kivonat: Artificial intelligence along with image processing technologies is the new driving force in the automotive industry. Nowadays automotive camera systems are key components of advanced driver assistance solutions (ADAS) allowing many comforts and active safety features such as lane keeping, road sign warning, emergency braking and semi-automated parking. Weather conditions causing visibility impairment significantly limit the applicability of driver-assistance cameras resulting in a higher risk to the driver’s safety. Machine learning algorithms processing camera signals can help to alleviate the problem of constrained visibility. However, there is a lack of training data because it is extremely difficult to take pictures of the same scene with and without haze. The overall goal of this project was to develop an application for generating synthetic training data by simulating photo-realistic fog with adjustable homogeneity on single images. Information that avers to be crucial to solve the problem such as the scene geometry, illumination of the scene (sunlight and artificial light sources), shadows etc. were extracted from pixel data. For instance, we used convolutional neural networks for estimating the depth map and we simulated the heterogeneity of the fog by using simplex noise moreover, we improved the physical model which allowed us to deal with shadows in a physically accurate way. Our algorithm can be considered efficient since the processing speed reaches 15 fps on PC, CPU only.