Summary
In recent years, the simultaneous positioning and mapping (SLAM) method dominated by visual sensors has received extensive attention and research. However, due to the real-time and accuracy of positioning, a certain visual calculation and map refresh frequency must be guaranteed, and the demand for computing and storage is high. , causing platforms with limited resources to lose support with iterative updates of the project. Therefore, a monocular vision SLAM method for ARM isomorphic processors is studied. This method first constructs a scale-free initial map based on pure vision, and solves the visual scale factor by aligning the IMU measurement data with absolute scale; then proposes a fast tracking strategy to improve Running speed; finally, the scale of the back-end optimization problem is limited by the sliding window algorithm, and the edge push and data integration of historical data are performed in a timely manner, effectively avoiding the surge of local computing and storage under long-term running. Experiments on the TUM visual inertial navigation dataset show that compared with the ORB-SLAM3 method, the average computing speed of this method on the Raspberry Pi is increased by 69.29%, and the absolute trajectory error is lower. It supports real-time push and offline storage of map data. , is an effective method for the application of visual SLAM in embedded system platforms.
0 Introduction
Simultaneous positioning and map construction ( Simultaneous Localization and Mapping ,