室内动态场景下基于语义关联的视觉SLAM方法

Visual SLAM method based on semantic association in indoor dynamic scenes

针对视觉SLAM在动态场景下鲁棒性不足的问题,提出一种适用于动态场景下的视觉SLAM算法——SAD-SLAM。该算法首先使用GCNv2网络进行特征提取,以获取分布均匀的特征点集合,并加快提取速度。然后使用YOLOv8-seg语义分割网络完成场景内物体的检测,并对推理得到的物体按照是否具备自主运动能力进行划分。同时提出一种语义关联方法,通过对潜在动态物体进行2D和深度层面过滤,以确定潜在动态物体运动的可能性。最后,构建了含有语义信息的稠密3D点云地图,并避免了动态物体的干扰。算法使用TUM数据集及真实场景进行实验验证,结果表明,相较于ORB-SLAM3及其他相关的动态SLAM算法,SAD-SLAM在动态场景下具有更好的定位精度。

In order to improve the robustness of visual SLAM in dynamic scenes,this proposed a new visual SLAM algorithm called SAD-SLAM.This algorithm actively extracted features using the GCNv2 network to obtain a set of evenly distributed feature points and accelerate the extraction speed.Additionally,it detected objects within the scene using the YOLOv8-seg semantic segmentation network and classified them based on their ability to move autonomously.Furthermore,it used a semantic association method to filter potential dynamic objects at both the 2D and depth levels,determining their likelihood of movement.Finally,it constructed a dense 3D point cloud map containing semantic information,avoiding interference from dynamic objects.The effectiveness of this algorithm is demonstrated through experiments using the TUM dataset and real-world scenes.The results show that compared to ORB-SLAM3 and other related dynamic SLAM algorithms,SAD-SLAM achieves better positioning accuracy in dynamic scene.