图像语义分割是计算机视觉的重要研究领域,是场景理解的关键技术之一。在无人驾驶领域,通过对道路场景进行高质量的语义分割,可为自动驾驶汽车的安全行驶提供保障。首先从道路场景语义分割的定义出发,探讨了目前该领域面临的挑战;其次,将语义分割技术划分为传统的分割技术,传统与深度学习相结合的分割技术和基于深度学习的分割技术,重点介绍了基于深度学习的语义分割技术, 并按照强监督、弱监督、无监督三种不同的网络训练方式对其进行了阐述;然后总结与道路场景语义分割相关的数据集以及性能评价指标,并在此基础上进行对比,分析常见的图像语义分割方法的分割结果;最后,对道路场景语义分割技术面临的挑战以及未来的发展方向进行了展望。

Image semantic segmentation is an important research field of computer vision and also one of the key technologies for scene understanding. In the field of unmanned driving, high-quality semantic segmentation of road scenes provides a guarantee for the safe driving of autonomous vehicles. First, this paper starts with the definition of semantic segmentation of road scenes and discusses the current challenges in this field. Second, this paper divides the semantic segmentation technology into a traditional segmentation technology, a traditional segmentation technology combined with deep learning and a segmentation technology based on deep learning, focuses on the semantic segmentation technology based on deep learning, and elaborates it according to three different network training methods of strong supervision, weak supervision and unsupervison. Then, the datasets and performance evaluation indicators related to the semantic segmentation of road scenes are summarized and compared, and the segmentation results using the common image semantic segmentation methods are analyzed. Finally, the challenges faced by the road scene semantic segmentation technologies and the future development direction are prospected.