VI-NeRF-SLAM: A Real-time Visual-Inertial SLAM with NeRF Mapping

In numerous robotic and autonomous driving tasks, traditional visual SLAM algorithms estimate the camera's position in a scene through sparse feature points and express the map by estimating the depth of sparse point clouds. However,practical applications require SLAM to create dense maps in real-time, overcoming the sparsity and occlusion issues of point clouds. Furthermore,it is advantageous for SLAM map to possess an auto-completion capability, where the map can automatically infer and complete the remaining 20% when the camera observes only 80% of an object. Therefore, a more dense and intelligent map representation is needed. In this paper, we propose a Visual-Inertial SLAM with Neural Radiance Fields reconstruction to address the aforementioned challenges. We integrate traditional rule-based optimization with NeRF. This approach allows for the real-time update of NeRF local functions by rapidly estimating camera motion and sparse feature point depths to reconstruct 3D scenes. To achieve better camera poses and globally consistent map, we address the issue of IMU noise spikes resulting from rapid motion changes, along with handling pose adjustments due to loop closure fusion. Specifically, we employ a form of widening the static noise covariance to refit the dynamic noise covariance. During loop closure fusion, we treat the pose adjustment between pre- and post- loop closure as a spatiotemporal transformation, migrating NeRF parameters from pre- to post- to expedite loop closure adjustments in NeRF mapping. Moreover, we extend this method to scenarios with only grayscale images. By expanding the color channels of grayscale images and conducting linear spatial mapping, we can rapidly reconstruct 3D scenes with only grayscale images. We demonstrate the precision and speed advantages of our method in both RGB and grayscale scenes.