Optimization of the Image Preprocessing Pipeline for Autonomous Driving : Correlating image quality settings, metrics and methods to the performance of Machine Vision/Machine Learning functions for autonomous vehicles

Abstract: Image quality is a well-understood concept in human viewing applications, particularly multimedia, but it is also becoming more prevalent in the vehicle space. It is essential to work in this area to consider the safety of autonomous vehicles. The idea behind this master thesis is to work on machine vision techniques to work on high-contrast images. Producing images from digital cameras start with capturing light from a sensor, which is turned into human or machine-readable images by various processing techniques. However, these techniques significantly impact the performance of algorithms using this image as input. Therefore, a thorough analysis is required to define a pipeline that would produce optimal image output from raw image sensor data. This pipeline is called Image Signal Processor and can be useful for finding a good image quality metric. Further, extend to finding proper components to the pipeline and optimizing hyperparameters for these components. Image Signal Processing, aimed at the human vision, is over-represented in research and often developed in machine vision is motivated by these studies. However, the challenge is not the same, and measuring and understanding the difference is essential.