Design Of Semantic Visual SLAM System Combined With Deep Learning

SLAM is the core technology for robots localization and mapping,and plays an important role in most autonomous mobile devices.With the rapid development of the robot industry,the requirements for SLAM technology are gradually increasing,traditional visual SLAM systems have many areas to be improved,such as improving the positioning accuracy in complex environments and helping robots to complete intelligent tasks that reach the semantic level.Deep learning methods have great advantages in various visual tasks.Therefore,this paper combines deep learning methods to study how to improve the traditional visual SLAM system,and designs the system from three parts:visual odometry,pose optimization and mapping.In the design of the visual odometry,in order to reduce the impact of dynamic objects,this paper uses the optical flow method to detect dynamic feature points,while using YOLOv3 to detect the bounding box of the target object.All the feature points of the dynamic object are removed according to the position of the dynamic feature points and the bounding box.In the methods of pose optimization,first of all,the advantages of the graph optimization method are briefly analyzed,and then a loop closure detection method based on convolutional neural network is proposed,that is,using the convolutional neural network to extract the feature vector of the key frame,and compare the similarity between key frames,according to the similarity,determine whether a loop occurs.In the part of mapping,based on the 3D point cloud map,this paper uses super voxel segmentation and object detection to build an semantic object library,and completes the fusion of semantic information and point cloud map.The experimental results show that the visual odometry in this paper has high accuracy pose estimation results in both static and dynamic scenes,compared with the ORB-SLAM2 system,the improvement is more obvious in dynamic scene,the improvement can be up to 90%.Compared with the method of using the bag-of-words model with ORB features,the loop closure detection algorithm in this paper has higher accuracy,under the same recall rate,the accuracy rate can be increased by about 20%.The semantic map generated in this paper can clearly distinguish the objects in the environment,which is conducive to robots to complete more intelligent tasks.