Research On Indoor Simultaneous Localization And Semantic Mapping

With the development and application of robot technology,it is convenient for our life.Home robot technology generally uses visual Simultaneous Localization And Mapping(SLAM)to achieve positioning and establish navigation map.In practical application,how to realize the rapid and accurate positioning of visual SLAM system and how to construct a map with riched environmental information has become a hot issue in visual SLAM research.Under this demand,this paper combines optical flow method with key points to optimize the front-end tracking thread of visual SLAM,accelerate data processing speed,and combine deep learning and three-dimensional point cloud processing technology to obtain the semantic information of the environment,and construct an excellent visual semantic SLAM system.The feature point method visual SLAM system takes more time to extract feature points,in this paper improved Fast corner points are used as key points,and the rotation invariance of key points is obtained by combining gray-scale centroids.The image pyramid model is used to obtain the key point information of the different scales images to make the key points have scale invariance.The correct matching of key points between frames affects the final camera pose calculation results,so this paper uses the LK(Lucas-Kanade)optical flow method to perform key points matching between frames firstly,and initially remove the mismatch results according to the angle difference of the matching key points,and then combine it with the Random Sample Consensus(RANSAC)algorithm solves the inter-frame pose transformation matrix for accurate mismatch elimination.The correct matching results of the key points obtained by the above methods are used for multiple iterations of the Pn P(Perspective-n-Point)and RANSAC methods to accurately calculate the camera pose.Finally,new key frames detection method based on the number of key points and the camera pose is used to detect new key frames.Experiments show that the image processing speed of this method is about 40 FPS,which is about 1.4 times that of the original visual SLAM system while maintaining high positioning accuracy.In view of the problem that visual SLAM systems usually only use the low-dimensional geometric information of the environment to construct navigation maps with sparse landmarks and lack high-level perception of environmental information,this paper applies the deep learning network Deep Lab3+ to the visual SLAM system to perform semantic segmentation based on key frames,and combines the segmentation result with the3 D point cloud data to make the point cloud data have semantic information.Then,the hypervoxel clustering method is used to process the dense point cloud data to achieve the point cloud segmentation effect,and the point cloud semantic information is merged with the point cloud segmentation result to generate a three-dimensional point cloud map with semantic information to construct a semantic map.In experiments,the method of this paper can distinguish most indoor objects well,basically realizes the construction requirements of semantic map,and obtains high-level semantic information of indoor environment,laying a solid foundation for future robot navigation and better human-computer interaction experience.