Research On Key Technologies Of Monocular Visual SLAM Based On Deep Learning Methods

In recent years,with the continuous development of artificial intelligence technology,unmanned autonomous devices such as unmanned aerial vehicles and unmanned vehicles are widely used in daily life.As the key technology of unmanned equipment in autonomous positioning and path planning tasks,visual simultaneous localization and mapping technology(V-SLAM)is of great importance.Among various V-SLAM algorithms,the methods based on monocular camera is extremely appealing due to its simple structure and low manufacturing cost.However,the monocular SLAM system may cause practical problems in engineering applications,which can be summarized as cumulative positioning error and map scale drift.In case of problems above,the deep learning-based convolutional neural network(CNN)can be used for replacing some certain functional modules of traditional SLAM system in order to improve its working efficiency and navigation accuracy.In order to improve the environment perception ability of monocular cameras,this paper proposes a monocular depth estimation network algorithm based on deep learning.The framework of proposed network can be adjusted for accommodating different working situations.For computers with predominant computing power,this paper designs an "encoding-decoding" CNN based on the Densenet framework to predict the depth value and the surface normal map simultaneously.Once predicted,the surface normal images are used for optimizing the quality of depth maps.As for small embedded devices such as drones,another CNN structure is built based on the lightweight network structure Mobilenet-V2.While training the depth value prediction network,it also sets up prediction training for image semantic information and uses semantic information to guide the depth Optimization of predicted value.According to the imaging principle of the monocular camera,the RGB(Red,Green,Blue)three-channel coded image is restored to a dense point cloud image according to the depth estimation information of the scene;at the same time,the ORB-SLAM2 algorithm is used to obtain the pose transformation and scene of the image sequence Sparse point cloud map;organically integrate the above dense point cloud with the sparse point cloud information,and use the dense point cloud to correct the monocular scale drift defect of the sparse point cloud;finally,use the sparse point cloud to correct the structure of the dense point cloud image Distortion defect,so the two are spliced into a continuous scene structure robustness dense map.Point-cloud images have defects such as poor structural robustness,excessive storage space,and rough surface topography.This paper separately converts them into octree navigation maps,grid navigation maps and TSDF dense navigation maps which are fit for more extensive scenarios.Eventually,the kind of map for practical navigation can be determined by considering its own advantages and disadvantages.In concluded,the depth estimation neural network designed in this paper can obtain the dense depth of surrounding environment in real time.What’s more,the monocular visual odometry(VO)algorithm can quickly extract the information matrix of position and posture of input images,which can be used for accurately tracking the movement of the monocular camera.In addition,once achieving the depth estimation images,the scale drift errors existing in generated sparse map of current scene can be avoided.Then,the dense 3D scene reconstruction algorithm proposed in this paper combines sparse maps constructed by the VO algorithm with the correspond dense scene depth information predicted by CNN to reconstruct a 3D navigation map with richer and more accurate details.Through the above improvement work,the ability of unmanned equipment to perceive unfamiliar environments and carry out real-time path planning has been improved.