基于语义分割的视觉SLAM算法研究

Research on Visual SLAM Algorithm Based on Semantic Segmentation

目前大多数的视觉SLAM算法基于静态环境的假设,环境中的动态物体容易引起位姿估计的不准确。提出一种针对动态环境的改进算法。基于DS-SLAM方案进行改进,首先采用自适应阈值提取ORB特征点并通过改进四叉树算法将特征点均匀化;之后采用稀疏光流法跟踪角点的运动,同时结合Segment语义分割线程的结果分割动态物体;最后采用几何约束滤除动态点,保留高质量的特征点进行位姿估计,完成定位和建图功能。利用TUM数据集进行精度评测,相比于DS-SLAM算法,改进算法的实时性提升了9.02%。动态环境中相机位姿误差缩小了38.94%。通过提高特征点的质量,结合光流法和语义分割的优势,提升了机器人系统的定位精度和实时性。

At present,most visual SLAM algorithms are based on the assumption of static environment,in which dynamic objects tend to cause inaccurate pose estimation.This paper presents an improved algorithm for dynamic environment.Based on DS-SLAM,the ORB feature points are extracted by adaptive threshold and homogenized by improved quadtree.Then,the sparse optical flow method is used to track the motion of corner points,and the dynamic objects are segmented based on the results of Segment semantics.Finally,geometric constraints are used to filter out dynamic points,and high-quality feature points are reserved for pose estimation to complete positioning and mapping.Compared with DS-SLAM algorithm,the accuracy of the improved algorithm is evaluated by TUM data set,and the real-time performance of the improved algorithm is improved by 9.02%.The camera pose error in dynamic environment is reduced by 38.94%.In this paper,the positioning accuracy and real-time performance of the robot system are improved by improving the quality of feature points and combining the advantages of optical flow method and semantic segmentation.