激光雷达网络系统：交叉口移动目标检测与跟踪算法

需积分: 9 174 浏览量更新于2024-07-28 收藏 2.09MB PDF 举报

本文探讨的是"利用激光雷达网络检测和跟踪交叉路口的移动物体"（Detection and Tracking of Moving Objects Using a Network of Laser Scanners）。在前文中，研究人员已经开发了一个系统，通过一组水平激光扫描器监控交叉路口。然而，本文的核心在于介绍一个专门针对分布式激光扫描数据的算法，用于识别进入交叉口的每个移动物体，估计其大小，并跟踪其位置、速度和方向，尤其是在拥挤且部分被遮挡的情况下。该算法的独特之处在于，它首次尝试利用新颖的数据源进行移动物体的检测与跟踪，这带来了前所未有的挑战和机遇。目标是设计出一个能够处理大流量、高密度交通场景的系统，具有对部分观测数据的鲁棒性，以及在复杂环境下正确处理数据关联的能力。实验结果显示，该算法在实际交叉路口的测试中表现出色，对于1063个穿越的物体，有988个得到了从入口到出口的完美追踪，追踪准确率高达92.9%。尽管取得了显著的成果，但该系统也存在局限性，比如可能受环境因素如天气或光照条件的影响，以及对于高速移动物体的识别能力。未来的研究将集中在提高算法的实时性和准确性，优化数据处理以减少误报和漏报，以及进一步探索如何处理动态环境中更复杂的物体行为。这篇文章为智能交通系统领域提供了重要的技术突破，展示了如何利用激光雷达网络有效地进行交叉路口的动态物体监测，这对于智能交通管理和安全至关重要。随着技术的发展，这种基于激光雷达的移动物体检测和跟踪系统有望在未来的自动驾驶、交通流量管理等领域发挥更大的作用。

658 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 13, NO. 2, JUNE 2012

As has been discussed in a previous section, objects might

be partially obscured due to occlusions or range failures. For

example, when a car is only partially observed on one side, if

we do not assume any strong model on the car or we do not

even know that the object is a car, an estimation of its size and

center point could be quite unreliable. Partial observation could

greatly affect the reliability of feature parameter estimation and

could subsequently reduce accuracy in tracking moving objects.

Developing an object model to support robust feature parameter

extraction on partial observation data is key to this procedure,

as well as for building a tracking module.

In addition, a desirable detection result is that all moving

objects be successfully detected with a minimal number of false

alarms. When a moving object enters the area of laser coverage,

software can easily ﬁnd the object as long as it gives a reﬂection

to the laser beam. However, because an object might be simul-

taneously measured by different laser scanners and because the

contour points of an object might be spatially disconnected due

to occlusions or range failures, multiple alarms could arise from

a single moving object. To reduce multiple alarms and have

more complete knowledge for feature parameter estimation,

developing an algorithm of grouping the measurements from

different laser scanners into the same object is another key to

this procedure.

In the succeeding sections, we deﬁne an object model, we

address a method of grouping the measurements from different

laser scanners to detect moving objects in the environment

and extract their feature parameters, and we conclude with an

experimental result that veriﬁes the algorithm.

A. Object Model

Fig. 5 describes a special characteristic of laser measure-

ments. Suppose that a laser scanner performs counterclockwise

scanning, and the horizontal contour of a car is measured by a

sequence of laser points from s to e [see Fig. 5(a)]. Simplifying

the shape of a car using a rectangle, edges that represent two

vertical sides of the car can be detected through a corner

detector and a line ﬁtting on the laser points. A directional

vector u

that is associated with each edge is deﬁned according

to the scanning order of laser points, e.g., from a point measured

later to one measured earlier.

Let u denote a directional vector that is extracted from the

data of a single laser scanner after its alignment to a reference

frame. We found that no matter where a laser scanner is placed,

directional vectors u are equal if they are observations on the

same side of the object [see Fig. 5(b)]. Suppose that v

s, i =

1,...,4arethedirectional vectors deﬁned on each side of a

car, and they compose a counterclockwise loop. By matching u

with v

s, we can ﬁnd which side of the object is measured so

that the laser points that correspond to u are used to update the

estimates of that speciﬁc side. In this work, we call u a support

vector of the side v

Based on t he aforementioned considerations, an object model

is deﬁned in this work with the feature parameters shown in

Fig. 5(c), where the shape of an object is simpliﬁed using a

rectangular model. Developing a more accurate model for each

kind of object will be addressed in future work. In addition,

Fig. 5. Deﬁnition to an object model. (a) Measurement to a car from a single

laser scanner. (b) Measurements to a car from a network of laser scanners.

TABLE I

ARAMETERS IN AN OBJECT MODEL

a reliability item is deﬁned for each feature parameter for the

sake of partial observations (see Table I). Currently, reliabilities

are estimated with binary values, i.e., true =1or false =0.In

the case of a directional vector, the reliability denotes whether

the side has a support vector (reliable)ornot(unreliable).

In the case of a corner point, if both neighboring sides are

supported, the corner point is a (reliable) one; otherwise, it is

a guess through other feature parameters on the object model

(unreliable). In the case of dimensional size, reliability tells

whether the corresponding feature parameter represents a full

dimensional size (reliable) or perhaps a partial one (unreliable).

In the case of a center point, which cannot be directly observed,

the reliability denotes whether the coordinates are estimated

from other reliable feature parameters. A more detailed de-

scription of each parameter in an object model can be found

in Appendix A.

剩余15页未读，继续阅读

JFu付振宇

粉丝: 0

激光雷达网络系统：交叉口移动目标检测与跟踪算法

detection and tracking of point features

Detection and Tracking of Point Features

Real-time multiple vehicle detection and tracking from a moving vehicle.pdf

KLT 算法 detection and tracking of point features1991.pdf

A Contour-Based Moving Object Detection and Tracking

Multiple Obstacle Detection and Tracking using Stereo Vision

Target detection and tracking

People detection and tracking using RGB-D cameras for mobile robots

face detection and tracking with camshift

model based vehicle detection and tracking

最新资源