A Design Of Visual SLAM System Based On Sweeper

There are a lot of robot technologies that can be used for autonomous positioning,navigation,obstacle avoidance and mapping.For example,autonomous cruise of cars,navigation of drones and autonomous obstacle avoidance,as well as the construction and path planning of sweeping robots,no matter what kind of robots involve the above technologies,SLAM technology is indispensable to escape.Whether it is a robot or a driverless car,there are four core and fundamental problems:positioning technology,tracking technology,path planning technology and control technology.For the first three of the four,SLAM plays a central role.However,in the subsequent application of mobile devices,there are still problems such as too sparse point cloud maps and too slow running speed on small mobile devices Therefore,a dense mapping SLAM system based on deep neural network optimization algorithm is proposed.The main research work of this paper is as follows:1)Due to the problem that the point cloud map is too sparse and the tracking track of camera is not accurate enough in ORB-SLAM2 system,a semi-dense SLAM system based on a monocular camera is proposed.When a semi-dense map is built with a monocular camera,it is impossible that every pixel can be considered as a feature point to calculate their descriptors.In this paper,Gauss triangulation is used to determine the depth of image points,and a set of depth values with probability distribution is obtained by combining polar line search and block matching techniques.Finally,a depth filter is used to make the depth estimation converge to a stable value.Experimental results show that the proposed semi-dense SLAM system based on monocular camera can map a more dense point cloud,and the synchronization tracking and positioning accuracy of the system is 9.13%higher than that of ORB-SLAM2 system due to the semi-dense point cloud map.2)Because monocular camera needs to carry out triangulation and use depth filter to make the depth converge when calculating the depth,the amount of computation is too large,so the monocular camera cannot meet the demand of constructing dense point cloud map.Based on the above problems,this paper further proposes a dense SLAM system based on depth camera.Thanks to the depth camera is able to measure the depth information directly,the direct use of 3D point cloud processing open source library(PCL)to extract dense point cloud in the whole map.Then we use a statistical filter to remove isolated points,and use the voxel grid filter on large point cloud data of the camera for downsampling,which can reduce the number of point cloud,can also let the irregular point cloud data be smoother.The finally,different maps fragments add together,forming a dense point cloud map.It has been proved that this method can reconstruct dense point cloud map and control the average error range of synchronous positioning and tracking of mobile robot within 5 cm,satisfying the accuracy required by indoor mobile robots.3)When the ORB-SLAM2 system matches different ORB features,it is easy to mismatch two points that have similar appearance but actually not two corresponding points.Therefore,the algorithm optimization of SLAM system is carried out through deep neural network,and a learning scheme of neural network is proposed to replace the part of ORB feature detection in SLAM system.Geometric correspondence is generated according to the transformation of feature points in adjacent frames,which is used to make a visual odometer.The convolutional neural network(CNN)and recurrent neural network(RNN)were combined for training to detect the location of key points and generate corresponding descriptors.Using the rigid body transformation,the network is optimized by transforming the point from the source coordinate system to the reference system.The precision of frame-to-frame matching can be improved by neural network method.Without closed-loop detection,our system can improve the accuracy of the tracking pose(position and angle)by 49.16%compared to ORB-SLAM2.In addition,the SLAM system in this paper has better robustness.Compared with the ORB-SLAM2 system,it can deal with more environmental characteristics and is not easy to lose the target.