Design Of Visual-inertial Integrated Navigation Algorithm Based On Embedded CPU/GPU Heterogeneous Platform

Visual-inertial integrated navigation is a typical autonomous navigation technology.It can provide reliable navigation and positioning services for unmanned platforms under the condition that satellite navigation signals are not available.It has broad application prospects in many fields,especially in the military field.In this paper,the design of visual-inertial integrated navigation algorithm and its implementation technology based on embedded CPU/GPU heterogeneous platform are studied for the demand of unmanned platforms for autonomous navigation.The main research work and contributions of this paper are summarized as follows:(1)The visual-inertial integrated navigation algorithm based on Extended Kalman Filter is studied.Firstly,the kinematic model of the strapdown inertial navigation system based on MIMU is introduced,and the error state differential equation of the inertial navigation system is analyzed.Secondly,by augmenting the state variables of the inertial navigation system,the definitions of state variables and error state variables of the integrated navigation system are given,and the error state differential equation of the integrated navigation system is derived.Then,the relationship between pose increment errors of the carrier and error state variables of the integrated navigation system is studied,and a visual measurement model is constructed.Finally,the EKF-based filtering algorithm of the integrated navigation is studied to realize integrated navigation.The vehicle experiments show that the integrated navigation algorithm designed in this paper has higher navigation and positioning accuracy(better than 1.0% of the travel distance)and can run in real time.The effectiveness and robustness of the algorithm are verified by experiments.(2)The implementation technology of the integrated navigation algorithm based on the embedded CPU/GPU heterogeneous platform is studied.In the integrated navigation algorithm of this paper,the extraction and matching of SURF features are computationally intensive tasks.It would make the integrated navigation algorithm difficult to run in real time,when SURF is implemented based on CPU.Thanks to the high degree of parallelism,SURF algorithm is very suitable for parallel acceleration using GPU.In order to reduce the time overhead of feature extraction and matching as much as possible,based on the algorithm research,this paper designs the implementation scheme of the integrated navigation algorithm based on the embedded CPU/GPU heterogeneous platform and studies the specific implementation technology.Based on the scheme,the principle of SURF algorithm and its parallel design and implementation based on CUDA,the heterogeneous design and implementation of Visual Odometry,the heterogeneous design and implementation of integrated navigation algorithm are studied.Experiments show that the CUDA-SURF algorithm designed in this paper greatly reduces the time overhead of SURF feature extraction and matching,and improves the operational efficiency of the integrated navigation system.The embedded heterogeneous integrated navigation system implemented in this paper has higher precision,stronger robustness,and can run in real time.