Analysis of 3D Dynamic Urban Scenes Based on LiDAR Point Cloud Sequences

Abstract

In recent years, the spread of domestic, industrial and research robots and the robotic vehicles made 3D sensory information (such as LiDAR sensors) more and more widespread opposed to the conventional 2D image sensory information.A modern LiDAR sensor on board of a vehicle provides enough data to make us able to interpret the space surrounding the vehicle even at large scenes containing a whole street segment.In this paper, we present a method to reconstruct.The input of this procedure is a LiDAR sensor mounted on a car, which outputs a data stream covering more than 100 meters radius of space, collecting data at 15 Hz.The recording is in real environment on the streets of Budapest in real time, while the processing is offline, implemented on CPU keeping in mind the future implementation on GPUs to reach real time data processing.The data is provided by a Velodyne HDL-64E high performance LiDAR device.The aim is to segment several region classes (such as roads, building walls, vegetation) and to identify specified objects (such as people, vehicles, traffic signs) in the point clouds through a presegmentation step.To achieve this classification, we need several features such as the color and geometrical properties of the specified objects and their possible geometrical and physical interactions.Also, we need to take into account the time domain features calculated based on the LiDAR data stream.After this presegmentation step we are able to reconstruct building facades in 3D and to track the detected objects in the 3D space.Also, in the future, this processed data set can be registered against 2D images provided by conventional cameras to reproduce realistic, colored 3D virtual spaces.