| Simultaneous Localization And Mapping(SLAM)refers to a subject riding a specific sensor,and without prior environmental information,builds a threedimensional model of the environment during the movement and estimates its own movement.After years of development,SLAM has gradually shown its role in many fields such as autonomous driving,augmented reality,robot motion and positioning.Among them,the visual SLAM Based on feature point method has the advantages of convenient use and low sensor cost.These advantages make the visual SLAM Based on feature point method get extensive attention and have a good application prospect.In the traditional visual SLAM system based on the feature point method,all calculations are completed by the CPU,and the existing hardware resources(such as GPU)cannot be fully utilized for acceleration;secondly,the feature point extraction algorithms used by the system are artificially designed the image feature point extraction algorithm with strong assumptions,which causes the accuracy of path estimation to decrease greatly when the system is running in a single-textured environment,or even estimation cannot be performed.This paper conducts research on this issue,with the help of Deep Learning(DL)technology which has brought tremendous development to computer vision in recent years,to try to solve the keypoint extraction and inherent scale problems of traditional visual SLAM system based on feature point method.The main tasks as follows:Propose the SP-Flow SLAM system using self-supervised optical flow keypoint extractor.This paper uses self-supervised learning to build a keypoint extractor,which can extract the feature points required by visual SLAM.First,the basic network capable of extracting keypoints is trained by using convolution neural network and synthetic data;then,two consecutive frames of images and optical flow algorithm are used to train the basic network,a real-time inference network SP-Flow capable of extracting good pairing keypoints is obtained;finally,the SP-Flow network is used as the core module of the keypoint extractor and applied to the visual SLAM system based on the feature point method to obtain SP-Flow SLAM.The experimental results show that SP-Flow SLAM improves the path estimation accuracy of traditional visual SLAM based on the feature point method,and ensures the robustness of the system running in a low-texture environment to a certain extent.Propose the SPD-Flow SLAM system using self-supervised geometric information depth estimation module.In the case of using a monocular sensor,the visual SLAM based on the feature point method cannot correctly estimate the three-dimensional information in the two-dimensional image,so there will be inherent scale problems.Therefore,inherent scale problems will occur,resulting in the map generated by the system is difficult to use directly.In order to overcome the scale problem of the visual SLAM system running in the monocular mode,this paper uses self-supervised learning and geometric information to construct a depth estimation module on the basis of SP-Flow SLAM.First,a teacher network which can estimate the depth of the scene is trained by geometric information;then,the teacher network is simplified by knowledge distillation to get a smallscale student network,and the student network can complete the function of teacher network;finally,take the student network as the core part of the depth estimation module and integrate it into SP-Flow SLAM to obtain the SPD-Flow SLAM system.Experimental results show that the SPD-Flow SLAM system proposed in this paper retains the advantages of SP-Flow SLAM and improves the scale problem of SP-Flow SLAM in monocular mode. |
