物体级语义视觉SLAM研究综述

Survey of object-oriented semantic visual SLAM

中文关键词  视觉SLAM  数据关联  语义分割  物体级地图

英文关键词  visual SLAM  data association  semantic information  Semantic mapping

中文摘要

      视觉同时定位与地图构建(Visual simultaneous localization and mapping, VSLAM)是自主移动机器人、自动 驾驶、增强现实(AR)等领域的关键技术. 随着深度学习的发展, 准确高效的图像语义信息在VSLAM领域得到了广 泛的应用. 与传统SLAM相比, 语义VSLAM利用语义信息提升了定位精度和鲁棒性,并通过物体级重建提高了环境 感知能力, 成为当前VSLAM领域的研究热点. 本文对近年来优秀的物体级语义SLAM 工作进行了阐述归纳和对比 梳理, 总结了该领域的4个关键问题, 包括物体表达形式、物体初始化方法、融合语义信息的数据关联算法和融合物 体级语义信息的后端优化方法. 同时, 对代表性方法进行了优缺点分析. 最后, 在现有技术成果和研究基础上, 对物 体级语义VSLAM面临的挑战和未来研究方向进行了展望和分析. 当前物体级语义SLAM仍面临着物体关联不准 确、物体优化框架不完善等问题. 如何有效使用和维护语义地图以应用于决策规划等任务, 以及融合多源信息以丰 富视觉感知是未来的研究热点.

英文摘要

      Visual simultaneous localization and mapping (VSLAM) is a key technology for autonomous robots, autonomous navigation, and AR applications. With the development of deep learning, accurate and efficient semantic information has been widely used in VSLAM. Compared with traditional SLAM, semantic SLAM leverages semantic information to improve the accuracy and robustness of localization, and enhances environmental perception ability by object-level reconstruction, which has became the trend in VSLAM research. In this survey, we provide an overview of semantic SLAM techniques with state-of-the-art object SLAM systems. Four key issues of semantic SLAM are summarized, including object representation, object initialization methods, data association methods, and back-end optimization methods integrating semantic objects. The advantages and disadvantages of the comparison methods are provided. Finally, we propose the future work and challenges of object-level SLAM technology. Currently, semantic SLAM still faces problems such as inaccurate object association and an unified optimization framework has not yet been proposed. How to effectively use and maintain semantic maps for the application of decision and planning tasks, as well as integrate multi-source information to enrich visual perception, will be future research hotspots.