Deep Learning And Traditional Vision SLAM Based Monocular SLAM

Simultaneous Localization and Mapping(SLAM)plays a very important role in the robot's realization of environment and self-awareness.Research on this issue is of great theoretical and practical value for making the robot to be highly intelligent.SLAM methods requires the robot to establish environmental map based on the data acquired by its own sensors,and determine its position in the scene.Next time it need update the previous state and build a new state by the subsequently acquired data.Due to the unique advantages of the monocular vision sensor,monocular SLAM has received extensive attention from international scholars and has become an important research direction in the field of SLAM.Currently,monocular SLAM approaches can be divided into indirect method and direct method.The indirect method relies on the reprojection error of the sparse image feature points extracted to add the constraint to the state of the system,but the direct method is based on the assumption that the pixel gray value of a particular view spot is invariant in two images.Now monocular SLAM methods have got good experimental results in the laboratory,but in practical applications,because of the complexity of the environment,there are still many problems in the monocular SLAM,such as the initialization scale,the scale drift,the loop closure detection and so on.Resolving sub-problems in SLAM based on deep learning is currently a highly active research in the field of computer vision,and has obtained many results.At the same time,how to rely on deep learning to promote the development of SLAM technology is also the most cutting-edge research in SLAM.In this paper,we study the combination of deep learning and traditional visual SLAM to solve its current problems and improve the performance of algorithms in complex scenes.Based on this,an algorithm is designed,which transform the combination of the deep convolution scene depth prediction network FCRN and the traditional DSO SLAM algorithm into the optimal fusion of Information obtained from each other based on the optimization framework.The introduction of deep learning makes the SLAM algorithm designed in this paper have the ability of learning,so this is called Learning SLAM(LSLAM)algorithm.In the initial stage of Monocular SLAM,the algorithm solves the problem of inability to estimate the initial scale by introducing the depth information obtained by the depth prediction network;In the phase of the motion estimation,the predicted depth information is used to supervise and correct the scale,thus effectively reducing the scale drift in the system.In addition,the depth information of key point in key frame is initialized under the help of the predicted depth information so as to improve the effect of follow-up graph optimization.In the first part of this paper,the visual SLAM and its development based on deep learning are expounded,and the algorithm framework designed in this paper is explained.The second part introduces the principle of the core parts of traditional visual SLAM from two view of direct and indirect methods,and analyzes the hidden problems.The third part gives a detailed introduction to the classic deep network architecture for scene depth prediction based on convolutional neural network.In the fourth part,the various parts of LSLAM algorithm are described in detail,and the verification experiment of rationality are performed for each part.In the fifth part,the hardware platform of this paper—the robot “Jappeloup” and the software platform—ROS(robot operating system)are introduced,and then the performance of the algorithm designed in this paper is verified by experiments.In order to verify the performance of LSLAM algorithm,we design 3 experiments.In the first two experiments,four data sets were selected to perform comparison experiments with the traditional DSO SLAM algorithm and the RGB-D SLAM algorithm on the most essential robot positioning.In the third experiment,the absolute translation error is compared with the DSO algorithm on two common motion patterns(quadrilateral and circle)of robot in the laboratory.The fourth experiment carried out a comprehensive contrast experiment with DOS algorithm in the indoor corridor.The four experiments all show that the LSLAM algorithm designed in this paper improves the positioning accuracy of the robot.In the face of complicated scenes,the algorithm has a better performance,which proves the effectiveness of the proposed optimal fusion framework.