Deep learning methods for autonomous driving scene understanding tasks: A review

초록

Scene understanding is an imperative task in autonomous driving; requiring sensory data for contextual information mining and decision-making. In recent years, there has been a plethora of research conducted for two-dimensional (2D) visual scene understanding in the literature, educing promising results. Due to the development of efficient deep learning models and the availability of visual sensory data, the vehicles become aware of their surroundings by developing vehicular perception and analyzing vehicular contextual knowledge in real-world situations. This paper aims to review the recent state-of-the-art methods using deep learning technologies for vision tasks of autonomous driving applications. We split the scene understanding tasks into major vision applications including; object detection, semantic segmentation, instance segmentation, and panoptic segmentation. This review investigates the current scene understanding research in accordance with the basic frameworks, properties, achievements, and analyzes the performance in terms of advantages and limitations by state-of-the-art methods that primarily rely on recent deep learning architectures. It summarizes the available benchmark datasets, and evaluation criteria used for the visual localization tasks by the computer vision research community. Lastly, the review includes a discussion on challenges in the autonomous driving scene understanding research that even if recently encountered by researchers, still remains open to date to be addressed. © 2025 Elsevier Ltd

키워드