Real-Time and Accurate Segmentation of 3-D Point Clouds Based on Gaussian Process Regression

Abstract

In LIght Detection And Ranging (LIDAR)-based object detection, accurate object segmentation is of great importance, since segmentation is an essential preprocessing step for other perception tasks, such as classification and tracking. For segmenting objects, most of the previous methods have tried to eliminate the ground first, which typically incurs considerable overhead in computation and inaccuracy in object detection with point clouds gathered by using 3-D LIDARs. However, in many real-time applications, such as automated driving, segmentation should be performed within a specified time, because even a small delay in computation could result in vehicle collisions. In this paper, we propose a real-time and accurate object segmentation algorithm for 3-D point clouds, which does not carry out ground extraction as a first step. In the proposed algorithm, we generate candidate points of objects and find their borders based on the integrated structure of a 2-D grid and an undirected graph, which enables fast processing and yields an accurate segmentation result independent of ground extraction error. In order to enhance segmentation accuracy, we employ Gaussian process, which reduces over-segmentation that separates an object into multiple portions. We apply two types of Gaussian process models to alternately provide cues for merging adjacent over-segmented objects. Experimental results demonstrate that this paper achieves a real-time processing speed and higher segmentation accuracy than previous works in most evaluation metrics. With the application to tracking, we show that the enhanced segmentation accuracy increases the tracking accuracy by 11.4% even in the worst case.