Map Reconstruction with Indoor Glass Detection based on LiDAR Point Cloud Images

LiDAR is widely used for indoor navigation and mapping of mobile robots. However, LiDAR faces significant challenges in detecting transparent indoor objects such as glass due to the transmission and diffuse reflection of laser beams, resulting in a significant amount of laser data not returning to the LiDAR sensor. This can lead to gaps in map construction or erroneous navigation. Detecting and reconstructing transparent glass structures indoors can significantly improve the navigation efficiency and safety of indoor mobile robots. Current mainstream methods often rely on the fusion of LiDAR point cloud data with ultrasonic sensor data or 3D point cloud clustering techniques for glass detection. While these methods effectively detect indoor glass, they impose high hardware requirements on robots and suffer from poor real-time performance. To address this challenge, this paper proposes an innovative algorithm for indoor glass detection and reconstruction based on LiDAR point cloud images. The algorithm uses the YOLOv9 neural network model, which encodes 2D point cloud data into an image format as the network input to achieve accurate detection of indoor glass locations. The detection results are then used to reconstruct the glass on a raster map. Experimental validation in indoor environments shows that the proposed method reduces the relative error by 85% compared to the original mapping algorithm, highlighting its potential to improve the navigation performance of indoor mobile robots.