动态场景下基于视觉的SLAM技术研究

SLAM Based on Deep Learning and Optical Flow Constraints in Dynamic Scenes

摘要:

针对视觉同时定位与地图构建(SLAM)技术在动态环境中存在定位精度低、地图虚影等问题,提出了一种基于深度学习的动态SLAM算法。该算法利用网络参数少且目标识别率高的YOLOv8n改善系统的视觉前端,为视觉前端增加语义信息,提取动态区域特征点。然后采用LK光流法识别动态区域的动态特征点,剔除动态特征点并保留动态区域内的静态特征点,提高特征点利用率。此外,该算法通过增加地图构建线程,剔除YOLOv8n提取的动态物体点云,接收前端提取的语义信息,实现静态语义地图构建,消除由动态物体产生的虚影。实验结果显示,在动态环境下该算法与ORB-SLAM3相比,定位精度提升92.71%,与其他动态视觉SLAM算法相比,也有小幅度改善。

Abstract:

To address the problems of low localization accuracy and map vignetting in Visual Simultaneous Localization and Mapping (VSLAM) technology in dynamic environments, this paper proposes a dynamic SLAM algorithm based on deep learning. The proposed algorithm utilizes YOLOv8n, which has few network parameters and a high target recognition rate, to improve the visual front end of the system, add semantic information to the visual front end, and extract the dynamic region feature points. The LK optical flow method is then used to identify the dynamic feature points in the dynamic region, eliminate these dynamic feature points, and retain the static feature points in the dynamic region so as to improve the utilization rate of feature points. In addition, the proposed algorithm increases the map construction thread, eliminates the dynamic object point cloud extracted by YOLOv8n, receives the semantic information extracted by the front end, constructs a static semantic map, and eliminates the virtual shadow produced by dynamic objects. Experimental verification indicates that the proposed algorithm improves the localization accuracy in dynamic environments by 92.71% as compared to that of ORB-SLAM3. Further, it achieves a small improvement compared with other dynamic vision SLAM algorithms.