When applied to Unmanned Aerial Vehicles(UAVs),existing Simultaneous Localization and Mapping(SLAM)algorithms are constrained by several factors,notably the interference of dynamic outdoor objects,the limited computin...When applied to Unmanned Aerial Vehicles(UAVs),existing Simultaneous Localization and Mapping(SLAM)algorithms are constrained by several factors,notably the interference of dynamic outdoor objects,the limited computing performance of UAVs,and the holes caused by dynamic objects removal in the map.We proposed a new SLAM system for UAVs in dynamic environments to solve these problems based on ORB-SLAM2.We have improved the Pyramid Scene Parsing Network(PSPNet)using Depthwise Separable Convolution to reduce the model parameters.We also incorporated an auxiliary loss function to supervise the hidden layer to enhance accuracy.Then we used the improved PSPNet to detect whether there is a movable object in the scene.If there is a movable object,its feature points will be removed in the tracking thread,and the removed feature points will not participate in the pose estimation of the camera.In addition,we proposed a filling method based on Generative Adversarial Networks(GANs)for the holes caused by dynamic object removal in the map,which employs a new auxiliary descriptor to assist GANs in restoring static scenes based on semantic information.The proposed system is evaluated on the TUM dataset,and the results indicate that the proposed method performs better than DynaSLAM and DS-SLAM on the TUM dataset.We experimented on the Cityscapes dataset,the improved PSPNet achieving an Intersection Over Union(IOU)of 0.812.展开更多
文摘When applied to Unmanned Aerial Vehicles(UAVs),existing Simultaneous Localization and Mapping(SLAM)algorithms are constrained by several factors,notably the interference of dynamic outdoor objects,the limited computing performance of UAVs,and the holes caused by dynamic objects removal in the map.We proposed a new SLAM system for UAVs in dynamic environments to solve these problems based on ORB-SLAM2.We have improved the Pyramid Scene Parsing Network(PSPNet)using Depthwise Separable Convolution to reduce the model parameters.We also incorporated an auxiliary loss function to supervise the hidden layer to enhance accuracy.Then we used the improved PSPNet to detect whether there is a movable object in the scene.If there is a movable object,its feature points will be removed in the tracking thread,and the removed feature points will not participate in the pose estimation of the camera.In addition,we proposed a filling method based on Generative Adversarial Networks(GANs)for the holes caused by dynamic object removal in the map,which employs a new auxiliary descriptor to assist GANs in restoring static scenes based on semantic information.The proposed system is evaluated on the TUM dataset,and the results indicate that the proposed method performs better than DynaSLAM and DS-SLAM on the TUM dataset.We experimented on the Cityscapes dataset,the improved PSPNet achieving an Intersection Over Union(IOU)of 0.812.
