动态场景下基于实例分割和三维重建的多物体单目 SLAM

针对大多数SLAM系统在动态环境下相机位姿估计不准确与环境语义信息利用不充分的问题，提出一种基于实例分割的关键帧检测和贝叶斯动态特征概率传播的动态物体检测算法，并对环境中存在的静态物体三维重建，以此构建一个动态环境下的多物体单目SLAM系统。该系统对关键帧输入图像进行实例分割与特征提取，获取潜在运动物体特征点集合与静态物体特征点集合；利用非运动物体特征点集合获取帧间位姿变换，普通帧利用贝叶斯对动静态特征点进行概率传播，利用静态特征点集实现对相机位姿的精准估计；在关键帧中对静态物体进行联合数据关联，数据充足后进行多物体三维重建，构建多物体语义地图，最终实现多物体单目SLAM。本文在TUM与Boon公开数据集上的实验结果表明，在动态场景下，相较于ORB-SLAM2算法，绝对位姿误差的均方根误差平均降低54.1%和58.2%。

Multi-object monocular SLAM based on instance segmentation and 3D reconstruction in dynamic scene

To address the problems of inaccurate camera pose estimation and insufficient utilization of environmental semantic information in most SLAM systems in dynamic environments, proposes a dynamic object detection algorithm based on the instance segmentation, keyframe detection, and Bayesian dynamic feature probability propagation, and three-dimensional reconstruction of static objects in the environment. To construct a multi object monocular SLAM system in a dynamic environment, the system performs instance segmentation and feature extraction on key frame input images, which could obtain a set of potential moving object feature points and a set of static object feature points. A set of non-moving object feature points is used to obtain inter frame pose transformation, Bayesian probability propagation of dynamic and static feature points are utilized for ordinary frames, and a set of static feature points is used to achieve accurate estimation of camera pose. Joint data association is performed on static objects in key frames, and after sufficient data is available, multi object 3D reconstruction is performed to construct a multi object semantic map. Finally, multi object monocular SLAM is achieved. The experimental results on TUM and Boon public dataset show that in dynamic scenarios, compared to the ORB-SLAM2 algorithm, the RMSE of APE decreases by 54. 1% and 58. 2% on average.