SPI(Spherical Point Indicator) : Point Feature Image Generation for Point-wise Classification and Dynamic Object Removal

Abstract

3D point cloud map is generated by accumulating LiDAR sensor data scanned at various locations and times. During scanning, dynamic objects are scanned in different poses depending on location and time, which degrades the quality of the map and negatively affects the localization. In order to improve the quality of the 3D point cloud map and for long-term management, effective to create a 3D point cloud map using only static objects by removing objects including dynamic possibilities. In this paper, we propose a Spherical Point Indicator (SPI) that can classify each point and remove dynamic probability objects through a method of generating feature images from individual points of a three-dimensional point cloud in scan units collected from a LiDAR sensor. SPI generates a unique feature image of each point by using the distribution information and intensity information of neighboring points centered on each point. The generated images are used as inputs to the CNN network and classified. The SPI feature image generation is applicable to all 3D point clouds, and all the scanned points as a result of classification are individually classified according to categories. Our approach can classify all included 3D point clouds using only one scan, and only static objects can be detected by filtering the dynamic possibilities object categories from the classified results.