Real-time 3D camera pose estimation using SLAM and GPU parallel computing on Jetson Orin

Abstract

Estimating the position of a camera in 3D space has numerous important applications in 3D scanning systems and mobile robotics. In this article, the authors investigate a method for estimating the camera's position in 3D space using the simultaneous localisation and mapping (SLAM) algorithm, leveraging the computational power of the graphics processing unit (GPU) on the Jetson Orin embedded computer to develop real-time localisation applications. By implementing visual odometry algorithms on the compute unified architecture (CUDA) platform, experimental results demonstrated a significantly faster processing speed and higher accuracy compared to CPUbased methods, showing remarkable performance improvements in real-time scenarios. This underscores the potential for application in autonomous robotics and 3D scanning technology. The system efficiently processes feature extraction and matching using binary descriptors, optimising the workflow through parallel computations on the GPU. The research provides a detailed assessment of the accuracy and speed of the system, highlighting the advantages of utilising GPUs on embedded systems for 3D camera position estimation, thereby facilitating the design and fabrication of compact outdoor handheld 3D scanning devices that can still meet real-time requirements.