ORB-SLAM3 Front-End Acceleration System Based on ZYNQ Platform

ORB-SLAM3, as a highly robust real-time positioning algorithm, can further enhance the platform's perception of the environment. However, when deployed on small embedded platforms, there are limitations in size, weight, and power, which face the problem of insufficient computing power of embedded platforms to serve SLAM algorithms in real-time. Therefore, this paper is based on the ZYNQ platform to accelerate the most time-consuming feature extraction in the ORB-SLAM3 front-end, in order to achieve real-time positioning of small embedded platforms. Specifically, the AR0144 binocular camera module is used to collect 640*400 color images in real time, which are converted into grayscale and sent to the FPGA acceleration module to obtain homogenized feature points. After verification, the FPGA acceleration algorithm has a relatively small loss of accuracy, with an average error of about 3.6mm compared to the trajectory running on the PC. Meanwhile, our work achieves a certain balance between running speed and power consumption compared to PC platforms and ZYNQ ARM.