一种用于无人车野外环境的负障碍检测方法

越野环境

需积分: 35 68 浏览量 2020-03-04 21:14:00 上传评论 1 收藏 5.18MB PDF 举报

资源推荐

资源详情

资源评论

Research Article

A feature matching and fusion-based

positive obstacle detection algorithm

for field autonomous land vehicles

Tao Wu

, Huihai Cui

, Yan Li

, Wei Wang

, Daxue Lui

and Erke Shang

1,2

Abstract

Positive obstacles will cause damage to field robotics during traveling in field. Field autonomous land vehicle is a typical

field robotic. This article presents a feature matching and fusion-based algorithm to detect obstacles using LiDARs for field

autonomous land vehicles. There are three main contributions: (1) A novel setup method of compact LiDAR is intro-

duced. This method improved the LiDAR data density and reduced the blind region of the LiDAR sensor. (2) A math-

ematical model is deduced under this new setup method. The ideal scan line is generated by using the deduced

mathematical model. (3) Based on the proposed mathematical model, a feature matching and fusion (FMAF)-based

algorithm is presented in this article, which is employed to detect obstacles. Experimental results show that the per-

formance of the proposed algorithm is robust and stable, and the computing time is reduced by an order of two mag-

nitudes by comparing with other exited algorithms. This algorithm has been perfectly applied to our autonomous land

vehicle, which has won the champion in the challenge of Chinese “Overcome Danger 2014” ground unmanned vehicle.

Keywords

3-D LiDAR, field ALV, positive obstacle detection, FMF algorithm

Date received: 12 March 2016; accepted: 27 November 2016

Topic: Field Robotics

Topic Editor: Yangquan Chen

Introduction

A field autonomous land vehicle (ALV) is a typical field

robot that drives off-road. In order to keep itself safe, the

vehicle needs to detect both positive and negative obstacles

by its own sensors.

In the past decades, many literatures

have emerged to address the problem of obstacle

detection.

2,3

But heretofore, ALVs still could not work well

in a field unknown scene. In urban or indoor scene, strong

assumptions such as the existence of flat ground surface are

employed. Thus, the rough road surface and lack of highly

structured components in field scene brought external

difficulties for obstacles.

Positive obstacle detection is widely researched and

many distinguis hed literatures have emerged recently.

5–7

In general, different kinds of sensors, such as thermal

infrared camera, color camera, stereo, 2-D or 3-D LiDAR,

or their combination, are employed to detect the obstacles.

The thermal infrared camera is typically us ed to detect

living organisms, such as animals at night.

Color camera

is also used to detect specific obstacles, such as vehicle

Unmanned System Institution, College of Mechatronics and Automation,

National University of Defense Technology, Changsha, People’s Republic

of China

Autonomous Land Vehicle Research Center, Changsha, People’s

Republic of China

Corresponding author:

Erke Shang, Autonomous Land Vehicle Research Center, Deya Road,

Changsha 410073, People’s Republic of China.

Email: erke1984@qq.com

International Journa l of Advanced

Robotic Systems

March-April 2017: 1–20

ª The Author(s) 2017

DOI: 10.1177/1729881417692516

journals.sagepub.com/home/arx

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 3.0 License

(http://www.creativecommons.org/licenses/by/3.0/) which permits any use, reproduction and distribution of the work without

further permission provided the original work is attributed as s pecified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/

open-access-at-sage).

detection,

human detection,

or other special obstacle

detection.

In these approaches, the feature of the targets

is captured or learned beforehand, and then, small windows

are slipped around the whole image to find the potential

obstacle by comparing the targets’ features. In order to find

these obstacles more easily, some approaches are needed to

estimate the ground plane beforehand from single image

or stereo data.

Then, these obstacles under the estimated

ground plane can be found directly and easily. Unfortu-

nately, regular cameras cannot be used at night, and the

quality of stereo data is not always satisfactory. Moreover,

the camera also could be easily disturbed by various

illuminations.

Recently, LiDAR has been widely a pplied in obstacle

detection because it can accurately get the range of

information.

13,14

A typical way for obstacle detection is

the ground segmentation method by using LiDAR

sensor.

3,15,16

Larson et al.

first analyzed the off-road

terrain by using the point cloud to identify the p otential

hazards. A height–length–density terrain classifier was

proposed in the work of Morton and Olson ,

where some

prior methods were described to provide a unified

mechanism for obstacle detections. We also adopted some

similar works for obstacle detection as in the study of

Wang et al.

However,itwasveryhardtosegmentthe

ground in a field envir onment, since bumpiness on the

ground of surface in those cases is very common.

In order to deal with the obstacle detection for field

ALVs, a novel feature matching and fusion (FMF)-based

algorithm is presented i n this arti cle. A n ew type of set up

method for LiDAR is introduced firstly: Two compact 3-

D LiDARs are mounted on two sides of the vehicle (as

shown in Figure 1). There are two advantages of this new

setup: (1) It greatly reduces the blind region around the

vehicle “viewed” by LiDARs. (2) It greatly improves the

density of data generated by LiDARs. Then, a mathemat-

ical model o f the point d istribution o f a single scan line is

deduced to simulate the real one. Based on the proposed

mathematical model, an FMF-based algorithm is pre-

sented. By matching the simulated scan line with the real

scan line, potential obstac les are detecte d in a 2-D para-

meter space. Experimental results show that the presented

algorithm is reliable and effective. The proposed algo-

rithm has been widely tested and finally applied to our

ALV, which has won t he champion of the Chinese chal-

lenge of Overcome Danger 2014 ground unmanned vehi-

cle (Figure 1).

The remainder of this article is organized as follows:

“Related works” section reviews some related works on

obstacle de tection. The details o f the presented nove l

setupmethodaredescribedin“Anewsetupmethodof

compact LiDAR” section. “A mathematical model to

simulate the point distribution under the new setup” sec-

tion describes the details of deducing a mathematical

model of the point distribution for a single scan line,

which is employed to generate an ideal one. In “The

FMF-based a lgorithm for obstacle detection” section, the

details of the proposed FMF-based algorithm are pro-

vided. In “Performance evaluation” section, lots of experi-

ments are carried out in different scenes. The article

concludes in the last section.

Related works

Obstacle detection is an essential task for field ALVs.

Therefore, it has received lots of attentions in past years.

As described in last section, many literatures have

emerged to address this problem. As a common sensor,

camera has been widely used for vehicle detection and

human detection. For example, Dalal

presented a histo-

gram of oriented gradients (HOG)-based a lgorithm to

detect obstacles by a single camera. In the study of

Dalal,

many obstacles, such as human, bicycle, motor-

cycle, ca r, bus, and other kinds of obstacles, are used to

test his algori thm , and a good score has bee n achieved.

The main idea of that algorithm can be described a s fol-

lows: (1) The HOG feature is captured and trained both

from positive sample sets and negative sample sets; (2)

once a new scene (captured by a camera) comes, slip

window on different sizes would traverse the whole image

to find whether there exists a ma tched feature, which is the

potential target.

Most frequently, in order to detect obstacles, data are

projected into an assumed or estimated ground plane first.

Hoiem et al.

presented a pe rspective method to put

objects into a ground plane. The algorithm “r eflects the

cyclical nature of the problem by all owing probabilistic

object hypotheses to refine geometry.” Therefore, obstacles

on the ground can be detected.

Figure 1. The new setup of LiDARs on our ALV: two compact

LiDARs (labeled by a red circle) are mounted on the two sides of

the vehicle top with a same fixed angle. This ALV won the

champion of the challenge. An HDL-64 LiDAR (labeled by a blue

circle) is upright equipped on the middle of the vehicle top for

comparison. ALV: autonomous land vehicle.

2 International Journal of Advanced Robotic Systems

Stereo camera is always employed to estimate the

ground plane before detecting potential obstacles.

In the

study of Santana et al.,

a stereo-based obstacle detection

approach is presented using visual saliency. In this

approach, saliency map is generated from a single image,

and a 3-D point cloud is generated from the stereo process-

ing. Then, a ground plane is estimated from both the sal-

iency map and the 3-D point cloud map. An identical idea is

also mentioned in the article.

Murarka et al.

proposed

an algorithm for detecting obstacles in an urban setting

using stereo vision and motion cues. A global color seg-

mentation stereo method is also introduced to compare its

performance in detecting hazards against traditional local

correlation stereo methods.

Since the camera could be disturbed by various illumi-

nations, LiDAR sensor is chosen instead of camera to

estimate the ground plane before obstacle detection in

the study of Puntambekar.

Puntambekar

interpreted the

LiDAR data into the geometry of the terrain to build the

terrain model, and then, the data in each segment are ana-

lyzed to find whether there is an obstacle in this segment.

Moosmann et al.

used a graph-based approach to segment

ground and objects from 3-D LIDAR scans, using a novel

unified , generic criterion based on local convexity mea-

sures. Among these approaches, how to estimate the ground

is the key point to detect obstacles. Thus, segmentation of

3-D LIDAR point clouds is specially investigated, and a set

of segmentatio n methods for various types of 3-D point

clouds are presented in the study of Douillard et al.

In order to improve the performance of detection, many

algorithms are presented to fuse multi-cues from different

kinds of sensors. Manduchi et al.

proposed an obstacle

detection techniq ue for autonomous navigation in field

environments by using a single-line laser and a stereo cam-

era. At first, the slant of surface patches in front of the

vehicle is analyzed and estimated by using the stereo cam-

era. Then, LiDAR data are clustered into each distinct sur-

face segment, which is generated by the stereo s ystem

beforehand. The obstacles in each segment are detected

according to different slants. These mentioned approaches

may achieve ideal results in urban or indoor environment.

However, detecting obstacles in field environment (rough

road surface or even there is not a plane) is more difficult

than that in urban or indoor environment because there is

no flat ground surface assumption anymore. Therefore, the

ground plane-based algorithm for obstacle detection may

not give a good result in field environment.

In order to accurately estimate the rough terrain, an

algorithm that reconstructs a 3-D surface was introduced

by Hadsell et al.

The similar work was also carried out by

Lang et al.

and Vasudevan et al.

The main drawback of

these approaches is that its high computation burden cannot

meet the real-time requirement for ALVs.

Considering the problem of detection obstacle in field

environment, Kuthirummal et al.

presented a novel

approach that was applica ble for both structured and

unstructured environments. They explicitly detected scene

regions that were traversable instead of attempting to expli-

citly identify obstacles by using 2-D grid of cells. In their

approach, LiDAR and stereo data were considered as input

and were mapped into individual cells, and histograms of

elevations of the points in each cell were computed to find

potential obstacles. The similar procedure was mentioned

in the study of Larson et al.

A point cloud produced by a

3-D laser range finder was employed by Larson et al.

analyze the potential obstacles in the off-road terrain.

Nowadays, LiDAR becomes very popula r in obstacle

detections. Among these obstacle detection methods,

25,26

LiDAR data are first mapped onto a grid map, and the

position of maximal value, minimal value, and median

value is registered and marked in the whole grid map. Thus,

by comparing the registered value with adjacent grids, each

potential obstacle can be detected. However, the road sur-

face would be bumpy, and there would be swaying when

the vehicle is driven in field road. The height from the

ground ge ne rat ed by the ob sta cle would not b e distinct

enough, especially for these small obstacles. Theretofore,

these traditional grid map-based algorithms still underper-

form in the obstacle detection for field ALVs. To alleviate

this problem, several developed grid map-based algorithms

were proposed recently. For example, Montemerlo and

Thrun

introduced a multi-grid representation approach

to improve the performance of obstacle detection in field

environment. Several maps with different kinds of resolu-

tions are used in the study of Montemerlo and Thrun.

During its detecting process, a high-resolution map is used

for detecting near obstacles, and a low-resolution map is

used for far obstacles. Another key point of this approach is

that it needs to build a whole information solution map

before its detection. As mentioned, this information solu-

tion map is built by a simultaneous localization and map-

ping (SLAM) algorithm beforehand. Nevertheless, SLAM

itself is a hard task that has not been resolved well, espe-

cially in field environment.

In order to deal with this problem, a novel idea is intro-

duced in this article: During the detecting process, only the

comparison between a djace nt two scan points collect ed

from the same scan line is used to analyze the potential

obstacles. Compared with the traditional grid map-based

algorithm, this idea has two merits: Firstly, the comparison

of result from two contiguous scan points would not be

seriouslyimpactedbytheroad’s bumpiness; secondly,

since only the relationship between two contiguous scan

points is considered, the computation efficiency would be

greatly improved and the computing time would be seri-

ously reduced. The main and only sensor for obstacle

detection in our proposed algorithm is a pair of HDL-32

compact LiDAR. Another HDL-64 LiDAR and a color

camera are employed as comparative experiment to show

the detection results more directly. Therefore, the 64-line

3-D LiDAR and the color camera do not join the process of

obstacle detection. In order to put forward the novel idea, a

Wu et al. 3

剩余19页未读，继续阅读

评论收藏

内容反馈

erke1984

粉丝: 1
资源: 5

一种用于无人车野外环境的负障碍检测方法

902项目ros+qt界面对沟检测ROS_Com.tar.gz

无人驾驶关键技术分析三篇.docx

用户均衡与系统最优原则下交通分配模型的建立与分析

线结构光传感器的远距离障碍物感知研究。

电信设备-一种变形履带式小型移动系统.zip

探究GPS技术在矿区地质勘探工程测量中的应用.pdf

高中必修一至必修五单词表.doc

经典巡线机器人电源系统研究

GPS在工程测量中的应用 (1).pdf

博客中聚类算法（K-means、FCM、DBSCAN、DPC）的数据集（免积分）

机器学习期末复习题及答案

神经网络回归预测--气温数据集

Mathwork+Matlab+编程手册

Ollama软件windows安装包(版本0.3.10)

中文短信数据集-带标签

时间序列预测模型实战案例(Xgboost)(Python)(机器学习)包括时间序列预测和时间序列分类，点击即可运行！

亚博K210模型训练部署

Plecs电力电子仿真PLECS41.64 电力系统仿真软件免安装版本

shape_predictor_68_face_landmarks.zip

hugging face的models-openai-clip-vit-large-patch14文件夹

改进版的yolov5+双目测距

Stable-Diffusion WEBUI 简体中文语言包（2023.05.30更新）

Matlab 基于支持向量机(SVM)的数据回归预测 SVM回归

基于鲸鱼优化算法优化VMD参数试看效果代码(目标函数为样本熵)

mock_kaggle.csv

电机故障数据集.rar

一款操作微信的RPA工具

最新资源