3DVPS: A 3D Point Cloud-Based Visual Positioning System

Visual positioning is a critical function in applications such as navigation and extended reality experiences. Recently, deep learning technologies, especially classification, had been implemented on the positioning task. However, to acquire a comprehensive positioning dataset and produce a high-performance neural network model is challenging. In this article, we propose to solve the issue by projecting training images from auto-generated 3D point cloud maps. By utilizing branch convolutional neural network (B-CNN) model, the "zoom-in" equivalent property results in favorable positioning accuracy and successful real-time implementation.