A comprehensive study of fisheye image compression and perception for autonomous driving

Abstract

Abstract Fisheye cameras are widely used in various fields, including automotive contexts for $$360^{\circ }$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mn>360</mml:mn> <mml:mo>∘</mml:mo> </mml:msup> </mml:math> near-field vision around vehicles, as well as in photography, robotics, underwater imaging, and virtual reality. However, conventional image compression techniques do not take into account the specific characteristics of fisheye images, such as radial distortion and wide-angle field of view, especially when operating at low bit rates. This can lead to degradation of image quality and distortion of geometric features that are essential for computer vision (CV) applications such as object detection, semantic segmentation, and motion estimation. Recent studies have highlighted the impact of various noise factors on automotive camera sensors, the challenges of correcting radial lens distortion, and the effects of image compression artifacts on fisheye camera visual perception tasks. In this work, a comprehensive study of fisheye image compression and perception using deep learning-based techniques is conducted. It is demonstrated that deep learning-based techniques achieve better compression performance and perceptual quality than conventional techniques, particularly at low bitrates crucial for automotive applications.