Robot Global Relocation Algorithm Based on Deep Neural Network and 3D Point Cloud

When a mobile robot is restarted or “hijacked” due to an emergency situation, it cannot determine its position and orientation in a known map using its own sensors in places without a GPS signal, such as indoors or outdoors with tall buildings, trees, and other dense areas. In both of these cases, the robot will be unable to find its position in the known map, resulting in a relocation failure. To address the relocation problem of mobile robots, this paper proposes a mobile robot global relocation method that combines deep neural networks and 3D point clouds, based on SLAM (Simultaneous Localization and Mapping) technology. The main idea is to use spherical projection to process the point cloud data, then feed it into a neural network to extract descriptors for similarity measurement, and perform pose calculation to accomplish relocalization. The application of SLAM technology enables the robot to achieve simultaneous localization and map building based on sensor data, accurately determining its own position even in the absence of GPS signals. Extensive experiments have been conducted to validate the proposed method, demonstrating its effectiveness in reducing relocation time and memory costs, as well as achieving relatively high relocation accuracy. Therefore, this method not only enhances the robustness of the system but also improves the feasibility of the overall system.