Research On Semantic SLAM Technology Based On Convolutional Neural Network

Simultaneous Localization and Mapping(SLAM)technology is a key technology for robots to sense the surrounding environment.This technology relies on static and stable feature points for localization and mapping.However,the traditional visual SLAM technology only selects feature points based on the geometric feature in the image,and it is impossible to determine whether the selected object is in a stationary state.Therefore,the current visual SLAM technology is mostly applied in environments such as indoor scenes and closed scenes with few dynamic objects.In order to apply SLAM technology to outdoor scenes such as street scenes,a semantic SLAM system based on convolutional neural network is proposed.Based on ORB-SLAM2 algorithm,the system uses convolutional neural network to segment the image semantically to obtain the label of objects in the image,and uses the optical flow method to eliminate dynamic feature points on moving objects.It can ensure that the feature points used to build the map are static and stable.The main work is as follows: firstly,analyze the classic semantic segmentation network and propose the network structure of semantic segmentation.Secondly,aiming at the problem that segmentation network does not segment street scene datasets with high accuracy,an image depth gating layer is used to improve segmentation.Thirdly,an algorithm is designed that combines the above segmentation network and optical flow method to remove dynamic feature points by using the object label and motion state.Finally,in order to solve the problem of the inefficient storage of traditional point cloud maps,an octree map algorithm is incorporated to compress the point cloud information to realize the storage and real-time update in a dynamic environment.This paper evaluates and validates the semantic SLAM system on actual scenes and datasets,and the experimental results show that the proposed semantic SLAM system can construct correct and usable semantic maps in outdoor complex environments such as street scenes.At the same time,it also proves the feasibility of using semantic SLAM technology in outdoor complex environments.