没有合适的资源?快使用搜索试试~ 我知道了~
目标跟踪Moving Target Classification and Tracking from Real-time Vid...
需积分: 10 23 下载量 55 浏览量
2010-10-18
20:17:40
上传
评论 1
收藏 1.73MB PDF 举报
温馨提示
试读
7页
目标跟踪Moving Target Classification and Tracking from Real-time Video
资源推荐
资源详情
资源评论
Moving Target Classification and Tracking from Real-time Video
Alan J. Lipton Hironobu Fujiyoshi
The Robotics Institute
Carnegie Mellon University
5000 Forbes Ave
Pittsburgh, PA, 15213
email:
f
ajl
j
hironobu
j
raju
g
@cs.cmu.edu
URL: http://www.cs.cmu.edu/˜ vsam
Raju S. Patil
Abstract
This paper describes an end-to-end method for extract-
ing moving targets from a real-time video stream, classi-
fying them into predefined categories according to image-
based properties, and then robustly tracking them. Moving
targets are detected using the pixel wise difference between
consecutive image frames. A classification metric is ap-
plied these targets with a temporal consistency constraint
to classify them into three categories: human, vehicle or
background clutter. Once classified, targets are tracked by
a combinationof temporaldifferencingand templatematch-
ing.
The resulting system robustly identifies targets of inter-
est, rejects background clutter, and continually tracks over
large distances and periods of time despite occlusions, ap-
pearance changes and cessation of target motion.
1. Introduction
The increasing availabilityof video sensorsand highper-
formance video processing hardware opens up exciting pos-
sibilities for tackling many video understanding problems
[9]. It is important to develop robust real-time video under-
standing techniques which can process the large amounts of
data attainable. Central to many video understanding prob-
lems are the themes of target classification and tracking.
Historically, target classification has been performed on
single images or static imagery [12, 7]. More recently, how-
ever, video streams have been exploited for target detection
[6, 14, 10]. Many methods like these, are computationally
expensive and are inapplicable to real-time applications, or
require specialised hardware to operate in the real-time do-
main. However, methods such as Pfinder [14],
W
4
[6] and
Beymer et al [2] are designed to extract targets in real-time.
The philosophy behind these techniques is the segmen-
tation of an image, or video stream, into object Vs. non-
object regions. This is based on matching regions of inter-
est to reasonably detailed target models. Another require-
ment of these systems is, in general, to have a reasonably
large number of pixels on target. For both of these rea-
sons, these methods would, by themselves, be inadequate
in a general outdoor surveillance system, as there are many
different types of targets which could be important, and it is
often not possible to obtain a large number of pixels on tar-
get. A better approach is one inwhich classification is based
on simple rules which are largely independent of appear-
ance or 3D models. Consequently, the classification metric
which is explored in this paper, is based purely on a target’s
shape, and not on its image content.
Furthermore, the temporal component of video allows a
temporal consistency constraint [4] to be used in the classi-
fication approach. Multiplehypotheses of a target’s classifi-
cation can be maintained over time until the system is con-
fident that it can accurately classify the target. This allows
the system to disambiguate targets in the case of occlusions
or background clutter.
Many systems for target tracking are based on Kalman
filters but as pointed out by [8], they are of limited use be-
cause they are based on unimodal Gaussian densities and
hence cannot support simultaneous alternative motion hy-
potheses. A few other approaches have been devised, for
example, (a) Isard and Blake [8] present a new stochas-
tic algorithm for robust tracking which is superior to pre-
vious Kalman filter based approaches, and (b) Bregler [3]
presents a probabilistic decomposition of human dynamics
to learn and recognise human beings (or their gaits) in video
sequences.
This paper presents a much simpler method based on a
资源评论
zhangjingwei713
- 粉丝: 11
- 资源: 7
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- 2023-04-06-项目笔记 - 第一百十五阶段 - 4.4.2.113全局变量的作用域-113 -2024.04.26
- 2023-04-06-项目笔记 - 第一百十五阶段 - 4.4.2.113全局变量的作用域-113 -2024.04.26
- htmlzwbjq_downyi.com.zip
- 无头单向非循环链表的实现(Test.c)
- 无头单向非循环链表的实现(SList.c)
- 浏览器重定向插件更新文件
- SSA-BP麻雀算法优化BP神经网络多特征分类预测(Matlab实现完整源码和数据)
- 粒子群算法优化BP神经网络PSO-BP的MATLAB代码(数值预测)
- 基于Springboot的一起看书平台.zip
- 无头单向非循环链表的实现(SList.h)
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功