智能视频事件分析与理解：最新进展与应用

4星 · 超过85%的资源需积分: 10 29 浏览量更新于2024-07-31 收藏 7.66MB PDF 举报

"《智能视频事件分析与理解》是一本汇集了国际专家的研究成果的书籍，由Jianguo Zhang、Ling Shao、Lei Zhang和Graeme A. Jones共同编著，属于《计算智能研究》系列的第332卷。该书反映了近年来在智能视频事件分析领域的前沿进展，针对互联网和媒体技术快速发展的背景下，如何处理背景复杂性、遮挡和交互等挑战进行了深入探讨。书中首先从理解有意义事件的角度，引入了基于手势交互的双重视角，以澄清事件的概念。接着，它详细介绍了基于子空间技术的运动分割方法，通过结合图像强度边缘邻域的线索。书中还涵盖了当前关于人体动作描述和识别的最新技术，特别强调了三维时空特征的应用。对于在复杂场景中的对象定位和检测，提供了高效的方法，这在体育视频、家庭环境监控以及安全监控等领域具有重要意义。此外，作者探讨了运动分析技术在多种应用中的实践，如体育比赛、日常生活监控以及视频搜索和索引等。这本著作不仅为研究人员提供了丰富的未来研究方向，也为视觉监控、人机交互和相关领域的专业人士提供了一个实用的参考工具和手册。《智能视频事件分析与理解》一书不仅适合计算机视觉、模式识别、信息安全和人工智能等领域的研究生，也对那些希望在视频分析领域取得突破的专业人士具有极大的价值。同时，它还延续了系列中的其他卷，如软计算生物特征识别、工程中的计算智能、科学和技术中的模型推理，以及大脑中的信息路由等多个主题，展现了计算智能研究的广泛性和深度。" 这本书的出版，标志着在智能视频事件分析领域的研究已经取得了一定的成果，但仍面临着许多开放问题，未来的研究将继续聚焦于提高性能、减少误报和漏报，以及更好地适应各种复杂环境。通过阅读这本书，读者可以深入了解这一领域的最新理论和实践进展，为自己的研究或工作提供有力支持。

4R.-D.Vatavu

with regards to the feedback mechanisms that transform into events in the human-

computer dialogue. The main purpose of the chapter is thus to bring further knowl-

edge to the current understanding of gesture-based interfaces by the means of

meaningful events and by looking at the topic from two different but equally im-

portant perspectives.

2 Events for Spotting and Detecting Gestures

Techniques for spotting gestures in video sequences are being implemented by de-

tecting and monitoring (or tracking) custom events deﬁned using location, time and

posture criteria. The successful detection of one event triggers gesture recording

while detection of another starts the gesture recognizer that classiﬁes the recorded

motion. This section analyses various event types and their associated criteria as

well as the algorithms and techniques employed for segmenting gestures from

sequences of continuous motion. The focus will be primarily oriented towards

vision-based computing but other capture technologies will be mentioned when the

techniques and algorithms used for segmenting human motions are relevant to the

discussion. Jaimes and Sebe [45], Poppe [38], Moeslund et al. [33, 34] and Erol et

al. [12] provide extensive surveys on vision-based motion capture and analysis as

well as on multimodal human-computer interaction and they represent good starting

points for a general overview on the advances in these ﬁelds.

We identify four different event types that have been used extensively either

singly or in various combinations for segmenting gestures in video sequences:

• Location represents a powerful cue for detecting postures as well as for segment-

ing interesting gestures from continuous movements. Requiring that a gesture

starts or ends in a predeﬁned region or knowing/learning that some locations in

the scene are more likely to contain valid gestures leads to great reduction in

algorithmic complexity;

• Posture information allows marking gesture commands in a way that feels natural

and accessible for the users to perform and model cognitively: for example, a

given posture could mark the beginning of a gesture while another signals its

ending. Posture is a robust cue that gives both the user as well as the system the

certainty that a gesture command is being entered: the system is able to ﬁlter

out the majority of movements while being interested only in the actual gesture

commands. Also, users are creating themselves a mental model for the interaction

process: commands are issued only if speciﬁed postures are being executed in a

similar manner to how click-like events work in standard WIMP interfaces;

• Tap and touch events can be detected by touch-sensitive materials as well as by

video cameras (most horizontal interactive surfaces use IR video cameras in order

to detect touch events on the tabletop). A tap or a touch is clearly perceived as a

marking event from both the system as well as the user’s perspective. Touching

clearly signiﬁes both intent as well as command during the interaction process;

• Custom-based events other than the above may be additionally used in order to

ease even further the gesture detection process. They usually relate to various

6R.-D.Vatavu

false positives. This is important in unknown scenarios where color or motion-

based reasoning may fail cause of complex backgrounds to analyze (such as

crowded scenes with multiple motion sources or objects that present the same

color as the human skin);

• Systems that employ general tracking algorithms usually require an initialization

stage during which the object to be tracked needs to be presented to the system.

A special location or a combination between location and posture are usually

employed in order to initialize such trackers;

• The segmentation process of a gesture is considerably facilitated if the start and

ending locations are known in advance. Segmenting gestures from continuous

movements is a difﬁcult task hence knowing where and when a gesture started

and when it ended by simple location event triggers is deﬁnitely a plus on system

robustness.

Location-based events are triggered whenever motion, color or other features are

identiﬁed in such predeﬁned regions of interest. They may specify when the ges-

ture started, when it ended or whether it passed through some intermediate stage

towards completion. If the scenario is known and the location ﬁxed and static (for

example when working at a desk with the video camera monitoring the tabletop as

Figure 3 illustrates), the regions may be simply deﬁned during ﬁrst-time installa-

tion and calibration of the system. For example, Vatavu and Pentiuc [50] use such

location-based events in order to detect and track hand postures only when hands

are positioned in predeﬁned regions on the surface of the interactive coffee table

prototype. As the scenario is a static one (or unlikely to change much since installa-

tion), the calibration procedure represented by identifying and deﬁning the regions

of interest is promising to produce robust results.

However, if the scenario is dynamic and likely to change over time (e.g. what

a mobile robot will see during its travel), the common solution is to use a robust

technique that would identify some aspect of the human body and then monitor

Fig. 3 Static scenario corresponding to a ﬁxed camera installation for which regions of in-

terest were deﬁned during system setup and calibration. Left: processing can be reduced to

limited regions only and even hand-wise delimitation is possible. Right: special regions such

as the keyboard (1) or printer (4) areas could trigger special events.

The Understanding of Meaningful Events in Gesture-Based Interaction 7

regions of interest around it. The preferred choice in this case would be to detect

the operator’s face as robust solutions such as the Viola and Jones face detector do

exist for accomplishing this task [52]. Once the location of the operator face and,

implicitly, the location of his head are known, the positions of other body parts

are inferred using simple geometry constraints. Figure 4 illustrates this idea with

several regions dynamically computed around the user’s head at one arm-length

around his body. Even more, detecting the face brings in another advantage: as the

face region will expose skin-colored pixels, on-the-ﬂy calibration of color detectors

can be performed [7] which increases system robustness and adaptability to the

environment.

Fig. 4 The user’s face can be robustly detected and used subsequently for inferring the loca-

tions of other body parts. Following [7] and [32], special regions of interest may be deﬁned

around the bounding rectangle of the user’s face. When the hand is detected in one particular

region an event will be triggered signaling the start, continuation or ending of a gesture. A

wave gesture can be described for example as a series of 1-2-1-2-1 location events.

Cerlinca et al. [7] deﬁne and use such regionsof interest around the human body in

order to facilitate segmentation and recognition of free hand gestures. Their purpose

is to transmit commands to a mobile robot that may circulate in any environment.

Ten such regions are deﬁned around the operator’s face which is reliably detected

using the Viola and Jones classiﬁer. Skin color-based thresholding is further applied

to these regions only in order to detect the operator’s hands. Combinations of several

active regions (for which hands were detected inside) correspond to various gesture

commands such as move forward, backward, turn left, turn right, etc.

Marcel [32] performs a similar partitioning of the space around the user’s face

combined with anthropometry constraints in order to detect the location of the hand.

The scenario is very similar to the one illustrated in Figure 4. The user’s intention for

producing a gesture is detected by monitoring the active windows that are formed in

the body-face space. When skin-colored objects are detected within these predeﬁned

windows, assumption of the user hand is taken and a posture recognizer further

employed.

Iannizzotto et al. [19] use the same principle of location-based events for de-

tecting when a gesture command starts. The authors’ Graylevel VisualGlove system

does not track the whole hand but rather the thumb and index ﬁngers only in order to

simulate mouse operations. When both ﬁngers are detected inside a predeﬁned area

8R.-D.Vatavu

(the initialization stage), the tracking procedure begins and actions are being recog-

nized. Also, whenever tracking is lost, the initialization procedure must be repeated

with the ﬁngers positioned in the same predeﬁned regions.

The same initialization concept but for the entire hand this time is implemented

in the HandVU system of Kolsch et al. [23] which employs tracking of multiple

features (ﬂock of birds). Figure 5 illustrates the general idea behind systems that

employ tracking algorithms: the tracker starts processing when the hand is present

in a given location, a necessary condition to compute the values of the features that

will be monitored and followed in the succeeding video frames.

Fig. 5 Trackers usually require initialization which consists in presenting the system with the

object to be tracked. An initialization event can be triggered whenever the hand is present at

some particular location. The features required for the particular tracker are computed when

the hand-detected event is triggered.

Malik and Laszlo [31] use location criteria in order to ease the segmentation pro-

cess when detecting the user’s hands for their Visual Touchpad system. The location

information is employed indirectly this time and is implemented by using a black

colored pad on top of which the hands can be reliably detected due to the high

difference in contrast between skin and background colors.

Locations and regions of interest represent powerful event triggers, irrespectively

whether they are being used for initialization only or during the actual process of

gesture detection and acquisition. They are usually combined with posture events for

which predeﬁned hand postures are being required in order for the location events

to be conﬁrmed. Two different scenarios can thus be identiﬁed for such events that

employ location criteria: initialization-only and continuous tracking and monitoring.

2.2 Posture-Based Events

Posture events are triggered whenever a predeﬁned posture is detected. They have

been used in order to mark the start and stop timestamps of a motion gesture

剩余260页未读，继续阅读

cupfire

粉丝: 6
资源: 12

智能视频事件分析与理解：最新进展与应用

"深度学习视频分析系统开发与部署指南

"研华智慧影像产品分享：视频采集卡、监控软件、交通解决方案详解

HCIA-Intelligent Video H12-511学习资料全新升级V2.0

Review of Intelligent Video Surveillance Technology Research

智能视频分析（Intelligent Video Surveillance）

Advances in Intelligent Data Analysis

HCIA-Intelligent Video Surveillance V1.0实验手册.rar

HCIA-Intelligent Video Surveillance V1.0 培训文档.rar

HCIA-Intelligent Video Surveillance V1.0培训文档 .rar

HCIA-Intelligent Video Surveillance（H12-511）教材V1.0.zip

最新资源