Image and video quality in the automotive environment

Abstract

This thesis is concerned with the development of methodologies for image and video quality assessment for the automotive environment, and the use of those methodologies for evaluating the effect of image and video degradations in automotive vision systems. Image quality metrics are important tools for optimizing system design parameters associated with image acquisition, compression and transmission. While optimizing systems for perceptual quality is already a common element of consumer electronics devices, in the automotive environment Advanced Driver Assistance Systems (ADAS) incorporating machine vision applications such as automated pedestrian detection are becoming a more widely-used feature of vehicular vision systems and the quality requirements of such systems present an additional challenge. As such, automotive image quality must also be tuned for optimal machine vision performance. In this thesis, quality is considered from the perspective of both machine vision performance and perceptual quality. An evaluation of the effect of image degradations on pedestrian detection performance is first carried out. This study highlights the quality impact of different imaging system degradations, such as compression artifacts, on pedestrian detection performance. It is demonstrated that improvements in detection performance can be achieved by training detection algorithms on images with a wide variety of degradations. A full-reference objective image quality assessment algorithm based on Histograms of Oriented Gradients (HOGs) is also proposed that correlates closely with pedestrian detection performance on degraded video frames. A system for No-Reference distortion classification, suitable for realtime operation, is also proposed. The classification system, based on natural image statistics, is combined with a multi-classifier approach to pedestrian detection in order to increase pedestrian detection performance on degraded images. Furthermore, a new approach for predicting the subjective Quality of Experience (QoE) of fish-eye to rectilinear transformed images is proposed. Improved correlation with subjective human opinion is achieved by weighting local quality scores with saliency information. Finally, an automotive specific video quality database is presented consisting of 50 video sequences and associated human saliency data and mean opinion scores. The influence of packet loss on visual QoE for high bandwidth automotive networks is also considered. The results show that increasing the level of packet loss has almost no effect on visual attention, despite significant differences in the MOS scores with different levels of packet loss.