Research On Key Technologies Of Semantics SLAM Based On Vision

Semantic SLAM is based on Simultaneous Localization and Mapping(SLAM)technology,using artificial intelligence(AI)technology for semantic segmentation.It can achieve synchronous acquisition of scene geometry information and semantic information.It is of great significance to improve the intelligent autonomous navigation ability of robots or unmanned systems.In this thesis,we use visual sensor as an information source to further study the key technologies of semantic SLAM.The main works and creations are as the following:(1)The development history of visual SLAM over the past 30 years is reviewed,and the advantages and disadvantages of the methods used in each development stage are analyzed.The system framework of visual SLAM is introduced,and the basic mathematical models and mathematical principles involved in visual SLAM are studied.(2)For the RGB-D SLAM,there is too much redundant information in the construction of dense point cloud maps,and there is a problem of image sticking.A mapping method of SLAM based on a look-up table is proposed.First,we estimate the eight neighborhood motion direction of the image by segmenting the image,and then use the direction of motion to optimize the mapping when we build the scene map.Experimental results show that it can improve the accuracy of pose estimation,greatly improve the speed of map building,and effectively improve the quality of map.(3)Aiming at the poor adaptability of visual SLAM system in dynamic scenario at present.A method based on the combination of look-up table and optical flow method is proposed to remove the influence of dynamic targets.We use optical flow to detect dynamic targets in the scene,and then use the look-up table to map the scene.Experiments show that this method can quickly eliminate the moving objects in the scene,thus improving the performance of visual SLAM in dynamic scenes.(4)The object detection method based on deep learning is studied,and an improved SLAM method to remove the influence of dynamic targets is implemented by using Faster R-CNN network.Object detection is used to eliminate dynamic targets in the scene to improve the effect of visual SLAM.Experiments show that this method can identify people in the scene effectively and eliminate people when building maps.(5)The Flood Fill algorithm is studied and a single-target semantic SLAM method is implemented in combination with the object detection network.First,the semantic information is extracted through object detection,and then semantic segmentation is done by Flood Fill algorithm,and then the semantic map of the scene is constructed.Experiments show that this method can quickly and accurately mark semantic information in point cloud map for single target in the scene.(6)The semantic segmentation method based on deep learning is studied,and a semantic SLAM method is implemented using PSPNet network.First,we extract semantic information from scene using semantic segmentation network,and then construct semantic map of scene by visual SLAM.Experiments show that this method can effectively construct static maps of scenes,and can also quickly remove the moving object of the scene through semantic information.