Previous Abstract Return to Session D5 Next Abstract

Session D5: Indoor and Urban Navigation and Mapping

Smartphone HD Map Updates Using Monocular-Inertial ORB-SLAM3 and Gaussian Splatting
Rhea Joyce Zambra, Saurav Uprety, Raymond Lee, and Hongzhou Yang, Department of Geomatics Engineering, University of Calgary
Date/Time: Friday, Sep. 20, 9:20 a.m.

Peer Reviewed

Gaussian splatting has emerged as a state-of-the-art 3D representation technique due to its high-fidelity and fast rendering capabilities. While it has been successfully integrated into light detection and ranging (LiDAR) and depth-enabled simultaneous localization and mapping (SLAM) algorithms, its potential for accurate outdoor 3D mapping using smartphone data remains underexplored. Pre-built high definition (HD) maps are vital for autonomous vehicles but are costly to maintain, motivating research in decentralized, smartphone-enabled HD map update systems. Existing solutions lack direct 3D-to-3D point cloud comparison, which could offer more robust updates by bypassing segmentation-based object detection. In this paper, we present a novel post-processing pipeline that generates dense, accurate, and near-scale HD maps from smartphone data, enabling updates to existing LiDAR and multi-sensor generated base maps. Our approach uses monocular-inertial ORB-SLAM3 to recover a scaled camera trajectory, which uses loop-closure and keyframe selection to alleviate drift in the localization and point cloud reconstruction. The ORB-SLAM3 keyframes are then used to initialize a 3D Gaussian Splatting render of the scene, which densifies the point cloud using the images, and is then scaled by the monocular-inertial camera trajectory. The camera and IMU data are collected using an iPhone 14 Pro Max, at an outdoor loop at the University of Calgary that spans 158 meters. Both sensors are observed using the SensorLogger application, and the camera-IMU calibration is performed through Kalibr. This system results in a successful closed-loop 3D Gaussian render, producing a point cloud with 8.70% scale error and 0.493m root mean square (RMS) value for iterative closest point (ICP) when referenced to a LiDAR-IMU base map, showing the potential of smartphones in visual-inertial HD mapping. Additionally, the registration of a parked car demonstrates the system's capability for accurate map updates when aligned with a LiDAR-based reference map.

Previous Abstract Return to Session D5 Next Abstract