Research And Implementation Of Parallelized VI-SLAM Algorithm Based On GPU

At present,intelligent devices such as mobile robots and UAVs(Unmanned Aerial Vehicles)are more and more widely used.The involved SLAM(Simultaneous Localization and Mapping)algorithm is an important theoretical and technical support for robot or UAV positioning and navigation.It can use cameras,lidar Sensors such as IMU(Inertial Measurement Unit)sense the surrounding environment information,calculate the position and attitude of the carrier relative to the surrounding environment,and create an environment map.VI-SLAM(Visual-Industrial Simultaneous Localization and Mapping)uses two sensors,camera and IMU,to sense the surrounding environment information.Because the two sensors have complementary effects,VI-SLAM algorithm has the advantages of high precision and better robustness.However,the calculation involved in VI-SLAM algorithm is complex and requires high computing power of hardware.The traditional VI-SLAM algorithm runs well on desktop devices and often loses real-time performance when runs on embedded devices.For UAVs and robots using embedded devices as control units,it is of great significance to study the methods to improve the portability of VI-SLAM algorithm on embedded devices.Some embedded devices are equipped with GPU modules.With the help of the parallel computing power of GPU,the computing power of such embedded devices can be effectively improved and the computing burden of CPU on the devices can be reduced,which makes it possible to run VI-SLAM algorithm in real time on such devices.This paper proposes a parallelization scheme of VISLAM algorithm based on GPU.Based on VINS-Mono algorithm,GPU parallel technology is used to parallelize it.Firstly,the back-end nonlinear least squares optimization module of VINSMono is parallelized.When solving the nonlinear least squares optimization problem,aiming at the time-consuming problem of solving and constructing the normal equation,this paper realizes the parallel construction of the normal equation based on GPU parallel computing technology,and realizes the parallel solution of the normal equation based on the parallel Cholesky decomposition algorithm.Secondly,in VINS-Mono,in order to control the computational complexity of the back-end optimization algorithm,when adding a new state quantity to be optimized,the old state quantity in the nonlinear optimization problem is marginalized based on Shure complement theory.In this process,matrix addition,subtraction and multiplication operations will be involved.This paper realizes the parallelization of these operations based on GPU parallel computing technology.In addition,the generalized inverse of the matrix needs to be solved in the process of marginalization.Based on the parallel SVD(Singular Value Decomposition)algorithm,this paper first decomposes the matrix,and then calculates the generalized inverse of the matrix on the two unitary matrices and a diagonal matrix,so as to realize the parallelization of the marginalization process in VINSMono algorithm.Finally,for the Harris corner extraction algorithm,FAST(Features From Accelerated Segment Test)corner extraction algorithm,LK(Lucas-Kanade)optical flow tracking algorithm and the calculation and matching algorithm of BRIEF(Binary Robust Independent Elementary Features)descriptor involved in VINS-Mono,this paper studies the characteristics of these algorithms,and transplants these algorithm programs in VINS-Mono to GPU for operation,so as to realize the parallelization of image feature point extraction and detection in VINS-Mono algorithm based on GPU.The experiment on the embedded device Jetson TX2 shows that the scheme proposed in this paper can run well on the heterogeneous computing model composed of CPU and GPU.Compared with VINS-Mono,the construction and solution speed of normal equation is the same,the speed of optical flow tracking and marginalization is significantly improved.Compared with other parallel SLAM algorithms based on GPU,the scheme proposed in this paper has a higher degree of parallelism,which can fully utilize the parallel computing power of GPU on the device,and make the allocation of hardware computing resources on embedded devices more reasonable.