一种面向动态环境下视觉同时定位和建图的图像预处理方法

An Image Pre-Processing Method for Visual Simultaneous Localization and Mapping in Dynamic Environments

摘要:

提出了一种用于动态环境下视觉同时定位和建图（SLAM）系统的图像预处理方法。该方法可以很容易地集成到现有视觉SLAM系统中，使其在高动态环境下能够稳定、准确和连续的工作。首先，提出了一种综合使用语义分割网络和光流估计网络的动态物体识别算法，鲁棒、准确地识别图像中潜在的动态物体。然后，为了检测与动态物体关联的阴影，提出了一种基于区域生长的阴影识别算法。最后，使用图像补全技术对剔除动态物体后的图像进行补全。将该图像预处理方法与双目ORB-SLAM2结合，并在KITTI数据集上进行了实验，实验表明所提出的图像预处理方法显著地提升了视觉SLAM系统的定位精度，并且图像预处理方法中的每一个模块都有着不可替代的作用。

Abstract:

This paper proposes an image preprocessing method for visual simultaneous localization and mapping (SLAM) systems in dynamic environments， which can be easily integrated into existing visual SLAM systems to enable stable， accurate， and continuous operation in highly dynamic environments. First， it proposes a dynamic object recognition algorithm that integrates the use of semantic segmentation networks and optical flow estimation networks to robustly and accurately identify potential dynamic objects in images. Then， to detect shadows associated with dynamic objects， it proposes a shadow recognition algorithm based on region growing. Afterwards， it uses image completion techniques to fill in the gaps left by the removal of dynamic objects from the images. Finally， it combines this image preprocessing method with stereo ORB-SLAM2 and conducts experiments on the KITTI dataset， which demonstrates that the proposed image preprocessing method significantly improves the positioning accuracy of visual SLAM systems. Each module in the image preprocessing method plays an irreplaceable role.