基于视觉惯性的动态环境下SLAM方法

SLAM method in dynamic environment based on visual inertia

针对传统视觉惯性SLAM在动态环境下则会出现鲁棒性差的问题,提出一种可用于室内的动态环境的视觉惯性SLAM方法.结合室内动态物体的特点,提出了一种基于先验假设的语义信息方法,利用Mask R-CNN现实潜在动态对象识别.为解决语义分割网络短时间内分割图片有限的问题,融合光流估计对未分割图像进行分割预测.最后,通过动态特征点过滤算法实现动态特征点与静态特征点的分离.在基于OpenLORIS数据集进行实验表明,该方法在高动态环境下能够有效提高SLAM系统的定位精度及鲁棒性.

To address the problem that traditional vision-inertial SLAM relies on static environments and reduces robustness in dynamic environments,this paper proposes a visual-inertial SLAM method specifically designed for indoor dynamic environments.We introduce a semantic information approach based on a priori assumptions by considering the characteristics of indoor dynamic objects.First,a semantic information method based on a prior assumption is proposed to recognize realistic latent dynamic objects using Mask R-CNN by combining the features of indoor dynamic objects.Second,we incorporate optical flow estimation to predict segmented images to address the time-consuming segmentation problem in a semantic segmentation network and the resulting finite number of segmented images.Finally,dynamic and static features are separated by a dynamic feature filtering algorithm.Experiments on the open vision-inertial dataset OpenLORIS show that this approach can effectively improve the localization accuracy and robustness of SLAM systems in highly dynamic environments.