视觉深度估计与点云建图研究进展

Progress of visual depth estimation and point cloud mapping

即时定位导航(SLAM)是无人驾驶和机器人实现自主移动的关键技术。目前广泛应用于SLAM技术中的激光雷达传感器存在成本高昂、激光点云空间分辨率低及难以获得精确的语义信息等一系列问题。视觉传感器可以有效避免上述问题,但是在深度预测和建图等方面需要更复杂的算法。近年来,随着处理器算力的提升、数据集的丰富和新机器视觉算法的出现,视觉深度预测和建图算法的精度和效率都有了明显提升。本文对现有视觉深度预测与视觉建图方法进行了总结,从视觉数据的采集和算法设计等方面进行分类阐述,最后针对应用场景和未来发展方向进行了分析。

Simultaneous location and mapping(SLAM)is a key technology for autonomous driving vehicles and robots to realize autonomous movement.The lidar currently used in SLAM technology presents a series of issues,including high cost,low spatial resolution of laser point clouds,and difficulty in obtaining accurate semantic information.In contrast,cameras can effectively avoid the above problems,but more complex algorithms are required in depth prediction and mapping.In recent years,with the progress in computing hardware,the continuous expansion of data sets,as well as the introduction of new computer vision algorithms,the accuracy and efficiency of visual depth prediction and mapping algorithms have been significantly improved.This article summarizes the existing methods of visual depth prediction and point cloud mapping and classifies the methodology based on data collection approach and algorithm design,then analyzes the application scenarios and prospect of visual depth estimation and point cloud mapping.