Upload
duy-dung
View
219
Download
0
Embed Size (px)
Citation preview
7/29/2019 Q. Cai, Human Motion Analysis a Review
1/13
Computer Vision and Image Understanding
Vol. 73, No. 3, March, pp. 428440, 1999
Article ID cviu.1998.0744, available online at http://www.idealibrary.com on
Human Motion Analysis: A Review
J. K. Aggarwal and Q. Cai
Computer and Vision Research Center, Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, Texas 78712
Received October 22, 1997; accepted September 25, 1998
Human motion analysis is receiving increasing attention from
omputer vision researchers. This interest is motivated by a wide
pectrum of applications, such as athletic performance analysis,
urveillance, manmachine interfaces, content-based image storage
and retrieval, and video conferencing. This paper gives an overview
of the various tasks involved in motion analysis of the human body.
We focus on three major areas related to interpreting human mo-ion: (1) motion analysis involving human body parts, (2) tracking a
moving human from a single view or multiple camera perspectives,
and (3) recognizing human activities from image sequences. Motion
analysis of human body parts involves the low-level segmentation
of the human body into segments connected by joints and recov-
rs the 3D structure of the human body using its 2D projections
over a sequence of images. Tracking human motion from a single
view or multiple perspectives focuses on higher-level processing, in
which moving humans are tracked without identifying their body
parts. After successfully matching the moving human image from
one frame to another in an image sequence, understanding the hu-
man movements or activities comes naturally, which leads to our
discussion of recognizing human activities. c 1999 Academic Press
1. INTRODUCTION
Human motion analysis is receiving increasing attention from
omputer vision researchers. This interest is motivated by ap-
plications over a wide spectrum of topics. For example, seg-
menting the parts of the human body in an image, tracking the
movement of joints over an image sequence, and recovering the
underlying 3D body structure are particularly useful for analy-
is of athletic performance, as well as medical diagnostics. The
apability to automatically monitor human activities using com-
puters in security-sensitive areas such as airports, borders, and
building lobbies is of great interest to the police and military.
With the development of digital libraries, the ability to automat-
cally interpret video sequences will save tremendous human
effort in sorting and retrieving images or video sequences us-
ng content-based queries. Other applications include building
manmachine user interfaces, video conferencing, etc. This pa-
per gives an overview of recent development in human motion
analysis from image sequences using a hierarchical approach.
In contrast to our previous review of motion estimation of a
igid body [1], this survey concentrates on motion analysis of the
human body, which is a nonrigid form. Our discussion c
three areas: (1) motion analysis of the human body stru
(2) tracking human motion without using the body parts
a single view or multiple perspectives, and (3) recognizi
man activities from image sequences. The relationship a
these three areas is depicted in Fig. 1. Motion analysis of t
man body usually involves low-level processing, such as
part segmentation, joint detection and identification, and tcovery of 3D structure from the 2D projections in an
sequence. Tracking moving individuals from a single vi
multiple perspectives involves applying visual features to
the presence of humans directly, i.e., without considering t
ometric structure of the body parts. Motion information s
position and velocity, along with intensity values, is emp
to establish matching between consecutive frames. Once f
correspondence between successive frames is solved, th
step is to understand the behavior of these features throu
the image sequence. Therefore, our discussion turns to rec
tion of human movements and activities.Two typical approaches to motion analysis of human
parts are reviewed, depending on whether a priori shape m
are used. Figure 2 lists a number of publications in thi
over the past years. In both model-based and nonmodel-
approaches, the representation of the human body evolve
stick figures to 2D contours to 3D volumes as the comp
of the model increases. The stick figure representation is
on the observation that human motion is essentially the m
ment of the supporting bones. The use of 2D contours is d
associated with the projection of the human figure in im
Volumetric models, such as generalized cones, elliptical
ders,and spheres, attemptto describe the details of a human
in 3D and, therefore, require more parameters for compu
Various levels of representations could be used for grap
animation with different resolutions [2].
With regard to the tracking of human motion without in
ing body parts, we differentiate the work based on wheth
subject is imaged at one time instant by a single camera o
multiple perspectives using different cameras. In both co
rations, the features to be tracked vary from points to 2D
and 3D volumes. There is always a trade-off between f
complexity and tracking efficiency. Lower-level features
as points, are easier to extract but relatively more diffic
428
7/29/2019 Q. Cai, Human Motion Analysis a Review
2/13
HUMAN MOTION ANALYSIS
FIG. 1. Relationship among the three areas of human motion analysis addressed in the paper.
rack than higher-level features such as blobs and 3D volumes.
Most of the work in this area is listed in Fig. 3.
To recognize human activities from an image sequence, two
ypes of approaches were addressed: approaches based on a
state-space model and those using the template matching tech-
nique. In the first case, the features used for recognition could be
points, lines, and 2D blobs. Methods using template matchingusually apply meshes of a subject image to identify a particular
movement. Figure 4 gives an overview of the research in this
area. In some of the publications, recognition is conducted us-
ng only parts of the human figure. Since these methods can be
FIG. 2. Past research on motion analysis of human body parts.
naturally extended to recognition of a whole body move
we also include them in our discussion.
This paper is an extension of the review in [3]. The or
zation of the paper is as follows: Section 2 reviews wo
motion analysis of the human body structure. Section 3 c
research on the higher-level tasks of tracking a human bod
whole. Section 4 extends the discussion to recognition of hactivity in image sequences based upon successfully tra
the features between consecutive frames. Finally, section 5
cludes the paper and delineates possible directions for futu
search.
7/29/2019 Q. Cai, Human Motion Analysis a Review
3/13
430 AGGARWAL AND CAI
FIG. 3. Past research on tracking of human motion without using body parts.
2. MOTION ANALYSIS OF HUMAN BODY PARTS
This section focuses on motion analysis of human body parts,
.e., approaches which involve 2D or 3D analysis of the hu-
man body structure throughout image sequences. Convention-
ally, human bodies are represented as stick figures, 2D contours,
or volumetric models [4]; therefore, body segments can be ap-
proximated as lines, 2D ribbons, and 3D volumes, accordingly.
Figures5,6,andand7showexamplesofthestickfigure,2Dcon-
our, and elliptical cylinder representations of the human body,
espectively. Human body motion is typically addressed by the
movement of the limbs and hands [58], such as the velocitiesof the hand or limb segments, or the angular velocity of various
body parts.
Two general strategies are used, depending upon whether in-
ormation about the object shape is employed in the motion
analysis, namely, model-based approaches and methods which
do not rely on a priori shape models. Both methodologies fol-
ow the general framework of: (1) feature extraction, (2) feature
orrespondence, and (3) high-level processing. The major differ-
FIG. 4. Past work on human activity recognition.
ence between the two methodologies is in establishing f
correspondence between consecutive frames. Methods
assume a priori shape models match the real images to a p
fined model. Feature correspondence is automatically ach
once matching between the real images and the model is
lished. When no a priori shape models are available, how
correspondence between successive frames is based upo
diction or estimation of features related to position, vel
shape, texture, and color. These two methodologies can a
combined at various levels to verify the matching betwee
secutive frames and, finally, to accomplish more complex
Since we have surveyed these methods previously in [4], wrestrict out discussion to very recent work.
2.1. Motion Analysis without a priori Shape Models
Most approaches to 2D or 3D interpretation of human
structure focus on motion estimation of the joints of bod
ments. When no a priori shape models are assumed, he
assumptions are usually used to establish the correspon
7/29/2019 Q. Cai, Human Motion Analysis a Review
4/13
HUMAN MOTION ANALYSIS
FIG. 5. A stick-figure human model (based on Chen and Lees work [11]).
of joints between successive frames. These assumptions im-
pose constraints on feature correspondence, decrease the search
space, and eventually, result in a unique match.
The simplest representation of a human body is the stick fig-
ure, which consists of line segments linked by joints. The motion
of joints provides the key to motion estimation and recognitionof the whole figure. This concept was initially considered by
Johansson [9], who marked joints as moving light displays
MLD). Along this vein, Rashid [10] attempted to recover a con-
nected human structure with projected MLD by assuming that
points belonging to the same object have higher correlations in
FIG.6. A 2D contour human model (similar to Leung and Yangs model [26]).
FIG. 7. A volumetric human model (derived from Hoggs work [1
projected positions and velocities. Later, Webb and Agg[11, 12]extendedto 3D structure recovery of Johansson-typ
ures in motion. They imposed the fixed axis assumption, w
assumes that the motion of each rigid object (or part of an
ulated object) is constrained so that its axis of rotation rem
fixed in direction. Therefore, the depth of the joints can be
mated from their 2D projections. Detailed review of [912
be found in [4]. All of these approaches inevitably demand a
degree of accuracy in extracting body segments and joints
segmentation problem is avoided by directly using MLD
implies their restrictions to human images with natural clo
Anotherwaytodescribethehumanbodyisusing2DcontIn such descriptions, the human body segments are anal
to 2D ribbons or blobs. For example, Shio and Sklansky
focused their work on 2D translational motion of human b
The blobs were grouped based on the magnitude and direct
the pixel velocity which were obtained using techniques si
to the optical flow method [16]. The velocity of each par
considered to converge to a global average value over se
frames. This average velocity corresponds to the motion o
whole human body and leads to identification of the whole
via region grouping of blobs with a similar smoothed vel
Kurakakeand Nevatia [17] attemptedto locatethe joint loca
in images of walking humans by establishing correspondbetween extracted ribbons. They assumed small motion bet
two consecutive frames, and feature correspondence was
ducted using various geometric constraints. Joints were fi
identified as the center of the area, where two ribbons ov
Recent work by Kakadiaris et al. [18, 19] focused on bod
decomposition and joint location from image sequences
moving subject using a physics-based framework. Unlike
where joints are located from shapes, here the joint is rev
only when the body segments connected to it involve mo
In the beginning, the subject image is assumed to be on
formable model. As the subject moves and new postures o
7/29/2019 Q. Cai, Human Motion Analysis a Review
5/13
432 AGGARWAL AND CAI
multiple new models are produced to replace the old ones, with
each of them representing an emerging subpart. Joints are de-
ermined based on the relative motion and shape of two moving
ubparts. These methods usually require small image motion
between successive frames which will not be a major concern as
he video sampling rate increases. Our main concern is still seg-
mentation under normal circumstances. Among the addressed
methods, Shio and Sklansky [15] relied on motion as the cue
or segmentation, while the latter two approaches are based onntensity or texture. To our best knowledge, robust segmentation
echnologies are yet to be developed. Since this problem is the
major obstacle to most of the work involving low-level process-
ng, we will not mention it repeatedly in later discussions.
Finally, we want to address the recent work by Rowley and
Rehg [20] which focuses on the segmentation of optical flow
fields of articulated objects. It is an extension to motion anal-
ysis of rigid objects using the expectation-maximization (EM)
algorithm [21]. Compared to [21], the major contribution of [20]
s to add kinematic motion constraints to each pixel data. The
trength of this work lies in its combination of motion segmen-ation and estimation in EM computation, i.e., segmentation is
accomplished in the E-step, and motion analysis in the M-step.
These two steps are computed iteratively in a forwardbackward
manner to minimize the overall energy function of the whole im-
age. The motion addressed in the paper is restricted to 2D affine
ransforms. We would expect to see its extension to 3D cases
under perspective projections.
2.2. Model-Based Approaches
In the above subsection, we examined several approaches to
motion analysis that do not require a priori shape models. In this
ype of approach, which is necessary when no a priori shape
models are available, it is typically more difficult to establish
eature correspondence between consecutive frames. Therefore,
most methods for the motion analysis of human body parts apply
predefined models for feature correspondence and body struc-
ure recovery. Our discussion will be based on the representa-
ions of various models.
Chen and Lee [22] recovered the 3D configuration of a moving
ubject according to its projected 2D image. Their model used 17
ine segments and 14 joints to represent the features of the head,
orso, hip, arms, and legs (shown in Fig. 5). Various constraints
were imposed for the basic analysis of the gait. The methodwas computationally expensive, as it searched through all pos-
ible combinations of 3D configurations, given the known 2D
projection, and required accurate extraction of 2D stick figures.
Bharatkumar et al. [7] also used stick figures to model the lower
imbs of the moving human body. They aimed at constructing
a general kinematic model for gait analysis in human walking.
Medial-axis transformations were applied to extract 2D stick
figures of the lower limbs. The body segment angle and joint
displacement were measured and smoothed from real image se-
quences, and then a common kinematic pattern was detected
or each walking cycle. A high correlation (>0.95) was found
between the real images and the model. The drawback t
model is that it is view-based and sensitive to changes of th
spective angle at which the images are captured. Hubers h
model [23] is a refined version of the stick figure repre
tion. Joints are connected by line segments with a certain d
of relaxation as virtual springs. Thus, this articulated
matic model behaves analogously to a mass-spring-dampe
tem. Motion and stereo measurements of joints are confine
three-dimensional spacecalledproximityspace (PS). The hhead serves as the starting point for tracking all PS loca
In the end, particular gestures were recognized based on t
states of the joints associated with the head, torso, and
The key to solving this problem requires the 3D positions
joints in an image sequence. A recent publication by Iw
[24]focused on real-time extraction of stick figures from m
ular thermal images. The height of the human image an
distance between the subject and the camera were precalib
Then the orientation of the upper half of the body was calc
as the principle axis of inertia of the human silhouette. S
cant points, such as the top of the head and the tips of theand feet, were heuristically located as the extreme points fa
from the center of the silhouette. Finally, major joints such
elbows and knees were estimated, based on the positions
detected points through genetic learning. The drawback
method is, again, that it is view-oriented and gesture-rest
For example, if the human arms are placed in front of the
there is no way to extract the finger tips using this metho
therefore, it fails to locate the elbow joints.
Niyogi and Adelson [8] pursued another route to est
the joint motion of human body segments. They first e
ined the spatial-temporal (XYT) braided pattern produc
the lower limb trajectories of a walking human and cond
gait analysis for coarse human recognition. Then the proj
of head movements in the spatial-temporal domain was lo
followed by the identification of other joint trajectories.
joint trajectories were then utilized to outline the contou
walking human, based on the observation that the human
is spatially contiguous. Finally, a more accurate gait an
was performed using the outlined 2D contour, which le
fine-level recognition of specific humans. The major co
we have with this work is how to obtain these XYT trajec
in real image sequences without attaching specific sensors
head and feet during image acquisition.In Akitas work [25], both stick figures and cone appro
tions were integrated and processed in a coarse-to-fine m
A key frame sequence of stick figures indicates the approx
order of the motion and spatial relationships between the
parts. A cone model is included to provide knowledge
rough shape of the body parts, whose 2D segments corre
to the counterparts of the stick figure model. The prelim
condition to this approach is to obtain these key frames
to body part segmentation and motion estimation. Perale
Torres also made use of both stick figure and volumetri
resentations in their work [26]. They introduced a pred
7/29/2019 Q. Cai, Human Motion Analysis a Review
6/13
HUMAN MOTION ANALYSIS
ibrary with two levels of biomechanical graphical models. Level
one is a stick figure tree with nodes for body segments and arcs
for joints. Level two is composed of descriptions of surface and
body segments using 3D graphical primitives. Both levels of the
model are applied in various matching stages. The experiments
were conducted in a relative restricted environments with one
subject and multiple calibrated cameras. The distance between
he camera and the subject was also known. We would like to
see extension to their methods when the imaging conditions aremuch more relaxed.
Leung and Yang [27] applied a 2D ribbon model to recognize
poses of a human performing gymnastic movements. They esti-
mated motion solely from the outline of the moving human sub-
ect. This indicates that the proposed framework is not restricted
by the small motion assumption. The system consists of two ma-
or processes: extraction of human outlines and interpretation of
human motion. In the first process, a sophisticated 2D ribbon
model (shown in Fig. 6) was applied to explore the structural
and shape relationships between the body parts and their asso-
ciated motion constraints. A spatial-temporal relaxation processwas proposed to determine if an extracted 2D ribbon belongs
o a part of the body or that of the background. In the end, a
description of the body parts and the appropriate body joints is
obtained, based on the structure and motion. Their work is one
of the most complete systems for human motion analysis, rang-
ng from low-level segmentation to the high level task of body
part labeling. It would be very interesting to see how the system
reacts to more general human motion sequences.
As we know, the major drawback of 2D models is their re-
striction to viewing camera angles. Applying a 3D volumetric
model alleviates this problem, but it requires more parameters
and more expensive computation during the matching process.In a pioneering work in model-based 3D human motion analysis
28], ORourke and Badler applied a very elaborate volumetric
model consisting of 24 rigid segments and 25 joints. The surface
of each segment is defined by a collection of overlapping sphere
primitives. A coordinate system is embedded in the segments
along with various motion constraints of human body parts. The
system consists of four main processes: prediction, simulation,
mage analysis, and parsing (shown in Fig. 8). First, the image
analysis phase accurately locates the body parts based on the
previous prediction results. After the range of the predicted 3D
ocation of the body parts becomes smaller, the parserfits theseocation-time relationships into certain linear functions. Then,
heprediction phase estimates the position of the parts in the next
frame using the determined linear functions. Finally, a simulator
embedded with extensive knowledge of the human body trans-
ates the prediction data into corresponding 3D regions, which
will be verified by the image analysis phase in the next loop.
Our question is whether the computation converges in the four
process loop when the initial prediction is random and whether
he use of an elaborate representation of a human figure is nec-
essary when we merely want to estimate the motion of body
parts.
FIG. 8. The overview of ORourke and Badlers system.
Elliptical cylinders are another commonly used repres
tions for modeling human forms in 3D volumes. Compar
ORourke and Badlers elaborate model, an elliptical cyl
could be well represented by fewer parameters: the length
axis and the major and minor axes of the ellipse cross sec
Hogg [29] and Rohr [30] used the cylinder model originat
Marr and Nishihara [31], in which thehuman body is repres
by 14 elliptical cylinders. The origin of the coordinate syst
located at the center of the torso. Both Hogg and Rohr inte
to generate 3D descriptions of a human walking by modHogg [29] presented a computer program (WALKER) w
attempted to recover the 3D structure of a walking person.
applied eigenvector line fitting to outline the human imag
then fitted the 2D projections into the 3D human model us
similar distance measure. In the same spirit, Wachter and N
[32] also attempted to establish the correspondence betw
3D human model connected by elliptical cones and a real i
sequence. Both edge and region information were incorpo
in determining body joints, their degrees of freedom (DOF
orientations to the camera by an iterated extended Kalman
A fair amount of work in human motion analysis has focon themotion of certain body parts.For example,Rehg etal
were interested in tracking the motion of self-occluded fi
by matching an image sequence of hand motion to a 3D fi
model. The two occluded fingers were rendered, again,
several cylinders. The major drawback to this work is tha
assumed an invariant visibility order of 2D templates of
two occluded fingers to the viewing camera. This assum
simplified the motion prediction and matching process bet
the 3D shape model and its projection in the image plane. H
ever, it also limits its application to image sequences cap
under normal conditions. Recent work by Goncalves et al
7/29/2019 Q. Cai, Human Motion Analysis a Review
7/13
434 AGGARWAL AND CAI
addressed the problem of motion estimation of a human arm in
3D using a calibrated camera. Both the upper and lower arm
were modeled as truncated circular cones, and the shoulder and
elbow joints were assumed to be spherical joints. They used per-
pective projection of a 3D arm model to fit the blurred image
of a real arm. Matching was conducted by recursively minimiz-
ng the error between the model projection and the real image
hrough dynamically adapting the size and orientation of the
model. Iterative searching for the global minimum might be aoncern for this approach.
Kakadiaris and Metaxas [35], on the other hand, consider the
problem of 3D model-based tracking of human body parts from
hree calibrated cameras placed in a mutually orthogonal con-
figuration. The 3D human model was extracted, based on its 2D
projections from the view of each camera using a physics-based
ramework [19] discussed in the previous section. Cameras are
witched for tracking, based on the visibility of the body parts
and the observability of their predicted motion. Correspondence
between the points on the occluding contours to points on the
3D shape model is established using concepts from projectivegeometry. The Kalman filter is used for prediction of the body
part motion. Since the authors in [19] claimed that their method
was general to multiple calibrated cameras, Gavrila and Davis
36, 37] extended it to four calibrated cameras for 3D human
body modeling. The class oftapered super-quadrics, including
ylinders, spheres, ellipsoids, and hyper-rectangles, were used
or the 3D graphical model with 22 DOFs. The general frame-
work for human pose recovery and tracking of body parts was
imilar to that proposed in ORourke and Badlers [28]. Match-
ng between the model view and the actual scene was based on
he similarity of arbitrary edge contours using a variant ofcham-
fer matching [38]. Finally, we address the recent work by Hunter
t al. [39], who applied mixture density modeling for 3D human
pose recovery from a real image sequence. The 3D human model
onsists of five ellipsoidal component shapes representing the
arms and the trunk with 14 DOFs. A modified EM algorithm was
employed to solve a constrained mixture estimation of the low-
evel segmented images. Bregler [40] also applied the mixture
of Gaussian estimation to 2D blobs in their low-level tracking
process. However, their ultimate goal was to recognize human
dynamics, which will be discussed in later sections.
All of these approaches encounter the problem of matching a
human image to its abstract representation by models of differ-ent complexity. The problem is itself nontrivial. The complexity
of the matching process is governed by the number of model pa-
ameters and the efficiency of human body segmentation. When
ewer model parameters are used, it is easier to match the feature
o the model, but it is more difficult to extract the feature. For
example, the stick figure is the simplest way to represent a hu-
man body, and thus it is relatively easier to fit the extracted lines
nto the corresponding body segments. However, extracting a
tick figure from real images needs more care than searching for
2D blobs or 3D volumes. In some of the approaches mentioned
above, the Kalman filter and the EM algorithm were used to ob-
tain good estimation and prediction of body part positions.
ever, these algorithms are computationally expensive and
guarantee convergence if the initial conditions are ill defi
3. TR ACKING HUMAN MOTION WITHOUT
USI NG BODY PARTS
In the previous section, we discussed human motion
ysis involving identification of joint-connected body parwe know, the tasks of labeling body segments and then lo
their connected joints are certainly nontrivial. On the other
applications such as surveillance requireonly tracking the
ment of the subject of interest. It will be computationally
efficient to track moving humans by directly focusing o
motion of the whole body rather than its parts. Knowin
the motion of individual body parts usually converges to t
the whole body in the long run, we could employ the m
of the whole body as a guide for predicting and estimati
movement of body parts; i.e., we could use the the human
as the coordinates for local motion estimation of its bodyIn this section, we will review methods for tracking a m
human without involving its body parts.
As is well known to computer vision researchers, track
equivalent to establishing correspondence of the image str
between consecutive frames based on features related to
tion, velocity, shape, texture, and color. Typically, the tra
process involves matching between images using pixels, p
lines, and blobs, based on their motion, shape, and other
information [41]. There are two general classes of corre
dence models, namely iconic models which use corre
templates and structural models which use image fe
[41]. Iconic models are generally suitable for any objec
only when the motion between two consecutive frames is
enough so that the object images in these frames are highl
related. Since our interest is in tracking moving humans,
retain a certain degree of nonrigidity, our literature rev
limited to the use of structural models.
Feature-based tracking typically starts with feature extra
followed by feature matching over a sequence of image
criteria for selecting a good feature are its brightness, co
size, and robustness to noise. To establish feature corre
dence between successive frames, well-defined constrain
usually imposed to find a unique match. There is again a trabetween feature complexity and tracking efficiency. Lowe
features such as points are easier to extract but relatively
difficult to track than higher-level features such as lines,
and polygons.
We will organize our discussion based on two scena
from a single camera view or from multiple perspectiv
both cases, various levels of features could be used to est
matching between successive frames. The major differe
that the features used for matching from multiple perspe
must be projected to the same spatial reference, while tra
from a single view does not have this requirement.
7/29/2019 Q. Cai, Human Motion Analysis a Review
8/13
HUMAN MOTION ANALYSIS
3.1. Tracking from a Single View
Most methods focus on tracking humans in image sequences
from a single camera view. Features used for tracking have
evolved from points to motion blobs.
Polana and Nelsons work [42] is a good example of point fea-
ure tracking. They considered that the movements of arms and
egs converge to that of the torso. In [42], each walking subject
mage was bounded by a rectangular box, and the centroid of thebounding box was used as the feature to track. Tracking was suc-
cessful even when the two subjects were occluded to each other
n the middle of the image sequence, as long as the velocity of the
centroids couldbe differentiated. The advantage of this approach
s its simplicity and the use of body motion to solve occlusion.
However, it only considers 2D translational motion and its track-
ng robustness could be further improved by incorporating other
features such as texture, color, and shape. Cai et al. [43] also
focused on tracking the movements of the whole human body
using a viewing camera with 2D translational movement. They
focused on dynamic recovery of still or changing backgroundmages. The image motion of the viewing camera was estimated
by matching the line segments of the background image. Then,
hree consecutive frames were adjusted into the same spatial ref-
erence with motion compensation. At the final stage, subjects
were tracked using the center of the bounding boxes and esti-
mated motion information. Compared to [42], their work is able
o handle a moving camera. However, it also shares the strengths
and weaknesses of the previous approach. Segen and Pingalis
44] people tracking system utilized the corner points of moving
contours as thefeatures for correspondence.These feature points
were matched in a forwardbackward manner between two suc-
cessive frames using a distance measure related to position andcurvature values of the points. The matching process implies
hat a certain degree of rigidity of the moving human body and
small motion between consecutive frames was assumed. Finally,
short-lived or partially overlapped trajectories were merged into
ong-lived paths. Compared to the one-trajectory approaches in
42, 43], their work relies on multiple trajectories and thus is
more robust if all matches are correctly resolved. However, the
robustness of corner points for matching is questionable, given
he condition of natural clothing and nonrigid human shapes.
Another commonly used feature is 2D blobs or meshes. In
Okawa and Hanatanis work [45], background pixels were votedas the most frequent value during the image sequence [15]. The
meshes belonging to a moving foot were then detected by in-
corporating low-pass filtering and masking. Finally, the distance
between the human foot and the viewing camera was computed
by comparing the relative foot position to the precalibrated CAD
floor model. Rossi and Bozzoli [46] also used moving blobs to
rack and count people crossing the field of view of a vertically
mounted camera. Occlusion of multiple subjects was avoided
due to the viewing angle of the camera, and tracking was per-
formed using position estimation duringthe periodwhen thesub-
ect enters the top and the bottom of the image. Intille and Bobick
[47, 48] solved their tracking problem by taking advanta
the knowledge of a so-called closed-world. A closed-w
is a space-time domain in which all possible objects pres
the image sequences are known. They illustrated their t
ing algorithm using the example of a football game, wher
background and the rules for play are known a priori. Ca
motion was removed by establishing homographic transf
between the football field model and its model using landm
in the field. The players were detected as moving blobstracking was performed by template matching the neighb
gion of the player image between consecutive frames. T
approaches are well justified, given the restricted environm
of their applications. We would not recommend using the
image sequences acquired under general conditions.
Pentland et al. [49, 50] explored the blob feature thorou
In their work, blobs were not restricted to regions due to m
and could be any homogeneous areas, such as color, tex
brightness, motion, shading, or a combination of these. S
tics such as mean and covariance were used to model the
features in both 2D and 3D. In [49], the feature vector of ais formulated as (x, y, Y,U, V), consisting of a spatial
and color (Y,U, V) information. A human body is constr
by blobs representing various body parts such as head, t
hands, and feet. Meanwhile, the surrounding scene is mode
a texture surface. Gaussian distributions were assumed for
models of the human body and the background scene. In th
pixels belong to the human body were assigned to different
part blobs using the log-likelihood measure. Later, Azarbay
and Pentland [50] recovered the 3D geometry of the blobs
a pair of 2D blob features via nonlinear modeling and r
sive estimation. The 3D geometry included the shape an
orientation of the 3D blobs along with the relative rotatiotranslation to the binocular camera. Tracking of the 2D
was inherent in the recovery process which iteratively sea
the equilibria of the nonlinear state space model across i
sequences. The strength of these approaches is their compr
sive feature selection, which could result in more robust pe
mance when the cost of computation is not a major conce
Another trend is to use color meshes [51] or clusters
for human motion tracking. In the work by Gunsel et al.
video objects such as moving humans were identified as g
of meshes with similar motion and color (YUV) informa
Nodes of the meshes were tracked in image sequences block matching. In particular, they focused on the applic
of content-based indexing of video streams. Heisele et al
initially adopted a clustering process to categorize obje
various prototypes based on color in RGB space and thei
responding image locations. These prototypes then serve
seeds for the parallel k-means clustering during which cen
of the clusters are implicitly tracked in subsequent images.
is also applied for tracking human faces [53, 54] and identi
nude pictures on the internet [55]. Since our interest is mo
motion analysis, we will not discuss these approaches in d
Overall, we believe color is a vital cue for segmentatio
7/29/2019 Q. Cai, Human Motion Analysis a Review
9/13
436 AGGARWAL AND CAI
racking of moving humans, as long as extra computation and
torage is not a major concern.
3.2. Tracking from Multiple Perspectives
The disadvantage of tracking human motion from a single
view is that the monitored area is relatively narrow due to the
imited field of view of a single camera. To increase the mon-
tored area, one strategy is to mount multiple cameras at vari-
ous locations of the area of interest. As long as the subject iswithin the area of interest, it will be imaged from at least one
of the perspectives. Tracking from multiple perspectives also
helps to solve ambiguities in matching when subject images are
occluded to each other from a certain viewing angle. In recent
years, more work on tracking of human motion from multiple
perspectives has emerged [5659]. Compared to tracking mov-
ng humans from a single view, establishing feature correspon-
dence between images captured from multiple perspectives is
more challenging because the features are recorded in different
patial coordinates. These features must be adjusted to the same
patial reference before matching is performed.Recent work by Cai and Aggarwal [59, 60] uses multiple
points belonging to the medial axis of the human upper body as
he feature to track. These points are sparsely sampled and as-
umed to be independent of each other, which preserves a certain
degree of nonrigidity of the human body. Location and average
ntensity of the feature points are used to find the most likely
match between two consecutive frames imaged from different
viewing angles. One feature point is assumed to match to its
orresponding epipolar line via motion estimation and precal-
bration of the cameras. Camera switching was automatically
mplemented, based on the position and velocity of the subject
elative to the viewing cameras. Experimental results of tracking
moving humans in indoor environments using a prototype sys-
em equipped with three cameras indicated robust performance
and a potential for real-time implementation. The strength of this
approach is its comprehensive framework, and relative lower
omputational cost given the degree of the problem complexity.
However, the approach still relies heavily on the accuracy of
he segmentation results, and more powerful and sophisticated
egmentation methods would improve its performance.
Sato et al. [56], on the other hand, treated a moving human as
a combination of various blobs of its body parts. All distributed
ameras are calibrated in a world coordinate system that corre-ponds to a CAD model of the surrounding indoor environment.
The blobs of body parts were matched over image sequences us-
ng their area, average brightness, and rough 3D position in the
world coordinates. The 3D position of a 2D blob was estimated,
based on height information, by measuring the distance between
he center of gravity of the blob and the floor. These small blobs
were then merged into an entire region of the human image us-
ng the spatial and motion parameters from several frames. Kelly
t al. [58, 57] adopted a similar strategy [56] to construct a 3D
environmental model. They introduce a feature called voxels,
which are sets of cubic volume elements containing information
such as which object belongs to this pixel and the hist
this object. The depth information of the voxel is also ob
using height estimation. Moving humans are tracked as a
of these voxels from the best angle of the viewing sy
A significant amount of effort is made in [57, 58] to con
a friendly graphical interface for browsing multiple ima
quences of a same event. Both methods require a CAD mo
the surrounding environment and tracking relies on findi
3D positions of blobs in the world coordinate system.
4. HUMAN ACTIVITY RECOGNITION
In this section, we move on to recognizing human move
or activities from image sequences. A large portion of lite
in this area is devoted to human facial motion recognitio
emotion detection, which fall into thecategory of elastic no
motion. In this paper, we are interested only in motion inv
a human body, i.e., human body motion as a form of artic
motion. The difference between articulated motion and e
motion is well addressed in [4]. There also exists a body oon human image recognition by applying geometric fea
such as 2D clusters [15, 60], profile projections [61], t
and color information [55, 53]. Since motion informatio
not incorporated in the recognition process, we will not in
them for further discussion.
For human activity or behavior recognition, most effort
used state-space approaches [62] to understand the huma
tion sequence [5, 63, 13, 14, 64, 65, 40]. An alternativ
use the template matching technique [42, 6668] to conv
image sequence into a static shape pattern and then com
it to the prestored prototypes during the recognition pr
The advantage of using the template matching techniqulower computational cost; however, it is usually more sen
to the variance of the movement duration. Approaches ba
state-space models, on the other hand, define each static p
as a state. These states are connected by certain probabi
Any motion sequence is considered as a tour going throug
ious states of these static poses. Joint probabilities are com
through these tours, and the maximum value is selected
criterion for classifying activities. In such a scenario, dura
motion is no longer an issue because each state can repea
visit itself. However, approaches using these methods u
apply intrinsic nonlinear models and do not have a closedsolution. As we know, nonlinear modeling typically re
searching for a global optimum in the training process,
requires expensive computing iterations. Meanwhile, sel
the proper number of states and dimension of the feature
to avoid underfitting or overfitting remains an issue.
following subsections, both template matching and state-
approaches [62] will be discussed.
4.1. Approaches Using Template Matching
The work [68] by Boyd and Little is perhaps the onl
which uses MLDs as the feature for recognition, bas
7/29/2019 Q. Cai, Human Motion Analysis a Review
10/13
HUMAN MOTION ANALYSIS
emplate matching. They assumed that a single pedestrian is
walking parallel to the camera and that the movement of the
subject is periodic. First, they computed n optical flow images
from n+ 1 frames and then chose m characteristics such as the
centroid of MLDs and moments from each optical flow images.
Thus the kth characteristic (1 km) in n frames forms a time-
varying sequence with a phasek. The differences between these
phases Fi =i m (1 km 1) are grouped into the fea-
ure vector for recognition. As can be seen from the procedure,his approach is restricted by the relative positions between the
walking subject and the viewing camera angle.
Perhaps the most commonly used feature for recognition of
human motion is 2D meshes. For example, Polana and Nelson
42] computed the optical flow fields [16] between consecu-
ive frames and divided each flow frame into a spatial grid in
both X and Y directions. The motion magnitude in each cell
s summed, forming a high-dimensional feature vector used for
recognition. To normalize the duration of the movement, they
assumed that human motion is periodic and divided the entire
sequence into a number of cycles of the activity. Motion ina single cycle was averaged throughout the number of cycles
and differentiated into a fixed number of temporal divisions.
Finally, activities were recognized using the nearest neighbor
algorithm. Recent work by Bobick and Davis [66] follows the
same vein, but extracts the motion feature differently. They in-
erpret human motion in an image sequence by using motion-
energy images (MEI) and motion-history images (MHI). The
motion images in a sequence are calculated via differencing be-
ween successive frames and then thresholded into binary val-
ues. These motion images are accumulated in time and form
MEI, which are binary images containing motion blobs. The
MEI are later enhanced into MHI, where each pixel value isproportional to the duration of motion at that position. Each
action consists of MEIs and MHIs in images sequences with
various viewing angles. Moment-based features are extracted
from MEIs and MHIs and employed for recognition using tem-
plate matching. Cui and Weng [67] intended to segment hand
mages with different gestures from cluttered backgrounds us-
ng a learning-based approach. They first used the motion cue
o fix an attention window containing the moving object. Then
hey searched the training set to find the closest match, which
served as the basis for predicting the pattern in the next frame.
This predicted pattern is later applied to the input image for ver-fication. If the pattern is verified, then a refinement process is
followed to obtain an accurate representation of the hand im-
age with any cluttered backgrounds excluded. As can be seen
from these approaches, the key to the recognition problem is
finding the vital and robust feature sets. Due to the nature of 2D
meshes, all the features addressed above are view-based. There
are several alternatives to the solution. One is to use multiple
views and incorporate features from different perspectives for
recognition. Another is to directly apply 3D features, such as
he trajectories of 3D joints. Once the features are extracted, any
good method for machine learning and classification, such as
FIG. 9. The basic structure of a hidden Markov model.
nearest neighbor, decision tree, and regression models, cou
employed.
4.2. State-Space Approaches
State space models have been widelyused to predict, estiand detect time series over a long period of time. One repr
tative model is the hidden Markov model (HMM), whic
probabilistic technique for the study of discrete times s
HMM has been used extensively in speech recognition, bu
recently has it been adopted for recognition of human m
sequences in computer vision [5]. Its model structure cou
summarized as a hidden Markov chain and a finite set of o
probability distributions [69]. The basic structure of an H
is shown in Fig. 9, where Si represents each state connect
probabilities to other states or its own, and y(t) is the observ
derived from each state. The main tool in HMM is the B
Welch (forwardbackward) algorithm for maximum likelestimation of the model parameters. Features for recognit
each state vary from points and lines to 2D blobs. We wi
ganize our discussions according to the complexity of fe
extraction.
Bobick [13] and Campbell [14] applied 2D or 3D Cart
tracking data sensed by MLDs of body joints for activity re
nition. To state more specifically, the trajectories of mu
part joints form a high-dimensional phase space, and this p
space or its subspace is employed as the feature to recog
Both researchers intend to convert the continuous motion
a set of discrete symbolic representations. The feature veceach frame in a motion sequence is portrayed as a point i
phase space belonging to a certain state. A typical gesture
fined as an ordered sequence of these states restricted by m
constraints. In Campbell [14], the learning/training proc
accomplished by fitting the unique curve of a gesture in
subset of the phase space with low-order polynomials. Ge
recognition is based on finding the maximum correlatio
tween the predictor and the current motion. Bobick [13], o
other hand, applied the k-means algorithm to find the cent
each cluster (or state) from the sample points of the traj
ries. Matching between trajectories of 3D MLDs to a ge
7/29/2019 Q. Cai, Human Motion Analysis a Review
11/13
438 AGGARWAL AND CAI
was accomplished through dynamic programming. The major
oncern to these approaches is how to reliably extract these 3D
rajectories from general video streams.
Goddards human movement recognition [63] focused on the
ow limb segments of the human stick figure. 2D projections of
he joints were used directly, without interpretation, as inputs,
and features for recognition were encoded by coarse orientation
and coarse angular speed of the line segmentsin the image plane.
Although Goddard [63] did not directly apply HMM in his workor discrimination of human gaits, he also considered a move-
ment as a composition ofevents linked by time intervals. Each
vent is independent of the movement. The links between these
vents are restrained by feature and temporal constraints. This
ype of structure, called a scenario, is easily extended by assem-
bling scenarios into composite scenarios. Matching real scene
kinematics to modeled movements involves visual motion fea-
ures in the scene, a priori events matched to the scenario, and
he temporal constraints on the sequence and time intervals.
Another commonly used feature for identifying human move-
mentsor activities is 2D blobs, which couldbe obtained from anyhomogeneous regions based on motion, color, texture, etc. The
work by Yamato et al. [5] is perhaps the first one on recognition
of human action in thiscategory. Mesh features of binary moving
human blobs are used as the low-level feature for learning and
ecognition. Learning was accomplished by training the HMMs
o generate symbolic patterns for each class. Optimization of the
model parameters is achieved using the BaumWelch algorithm.
Finally, recognition is based on the output of the given image
equence using forward calculation. Recent work by Starner and
Pentland [64] applied a similar method to recognition of Amer-
can sign language. Instead of directly using the low-level mesh
eature, they introduced the position of a hand blob, its angle of
axis of least inertia, and eccentricity of its bounding ellipse in
he feature vector. Brand et al. [65] applied the coupled HMMs
o recognize the actions by two associated blobs (for example,
he right- and left-hand blobs in motion). Instead of directly
omposing the two interacting process into a single HMM pro-
ess whose structural complexity is the product of that of each
ndividual process, they coupled the two HMMs in such a way
hat the structural complexity is merely twice that of each in-
dividual HMM. Breglers paper [40] presents a comprehensive
ramework for recognizing human movements using statistical
decomposition of human dynamics at various levels of abstrac-ions. The recognition process starts with low-level processing,
during which blobs are estimated as mixture of Gaussian mod-
els based on motion and color similarity, spatial proximity, and
prediction from the previous frame using EM algorithm. During
his process, blobs of various body parts are implicitly tracked
over image sequences. In the mid-level processing, blobs with
oherent motion are fitted into simple movements of dynamic
ystems. For example, a walking cycle is considered as compo-
ition of two simple movements, the one when the leg has the
group support and the one when the leg swings in the air. At the
highest level, HMMs are used as a mixture of these mid-level
dynamic systems to represent the complex movement. R
nition is completed by maximizing the posterior probabi
the HMM model from trained data given the input ima
quence. Most of these methods include low-level process
feature extraction. But again, all recognition methods rely
2D features are view-based.
5. CONCLUSION
We have presented an overview of past developments
man motion analysis. Our discussion has focused on three
tasks: (1) motion analysis of human body parts, (2) high
tracking of human motion from a single view or multipl
spectives, and (3) recognition of human movements or act
based on successfully tracking features over an image sequ
Motion analysis of the human body parts is essentially the
3D interpretation of the human body structure using the m
of the body parts over image sequences, and involves low
tasks such as body part segmentation, joint location, and
tion. Tracking human motion is performed at a higher le
which the parts of the human body are not explicitly iden
e.g., the human body is considered as a whole when establ
matches between consecutive frames. Our discussion is c
rized by whether the subject is imaged at one time instant
single view or from multiple perspectives. Tracking a sub
image sequences from multiple perspectives requires th
tures be projected into a common spatial reference. Th
of recognizing human activity in image sequences assume
feature tracking for recognition has been accomplished
types of techniques are addressed: state-space approach
template matching approaches. Template matching is sim
implement, but sensitive to noise and the time interval movements. State-space approaches, on the other hand,
come these drawbacks but usually involve complex ite
computation.
The key to successful execution of high-level tasks is to
lish feature correspondence between consecutive frames,
still remains a bottleneck in most of the approaches. Typ
constraints on human motion or human models are impo
order to decrease the ambiguity during the matching pr
However, these assumptions may not conform to general
ing conditions or may introduce other tractable issues, s
the difficulty in estimating model parameters from the reage data. Recognition of human motion is just in its in
and there exists a trade-off between computational cost an
tion duration accuracy for methodologies based on state
models and template matching. We are still looking forw
new techniques to improve the performance and, meanw
decrease the computational cost.
ACKNOWLEDGMENTS
The work was supported in part by the United States Army R
Office under Contracts DAAH04-95-I-0494 and DAAG55-98-1-0230
Texas Higher Education Coordinating Board Advanced Technology
7/29/2019 Q. Cai, Human Motion Analysis a Review
12/13
HUMAN MOTION ANALYSIS
95-ATP-442 and Advanced Research Project 97-ARP-275. Special thanks also
go to Ms. Debi Paxton and Mr. Shishir Shah for their generous help in editing
and commenting on the paper.
REFERENC ES
1. J. K. Aggarwal and N. Nandhakumar, On the computation of motion of
sequences of imagesA review, Proc. IEEE76(8), 1988, 917934.
2. S. Bandi and D. Thalmann, A configuration space approach for efficient
anomation of human figures, in Proc. of IEEE Computer Society Workshopon Motionof Non-Rigidand Articulated Objects, PuertoRico,1997,pp.38
45.
3. J. K. Aggarwal and Q. Cai, Human motion analysis: A review, in Proc. of
IEEE Computer Society Workshop on Motion of Non-Rigid and Articulated
Objects, Puerto Rico, 1997, pp. 90102.
4. J. K. Aggarwal, Q. Cai, W. Liao, and B. Sabata, Non-rigid motion analysis:
Articulated & elastic motion, CVGIP: Image Understanding, 1997.
5. J. Yamato, J. Ohya, and K. Ishii, Recognizing human action in time-
sequential imagesusingHiddenMarkov Model, in Proc. IEEE Conf. CVPR,
Champaign, IL, June 1992, pp. 379385.
6. W. Kinzel, Pedestrian recognition by modelling their shapes and move-
ments, in progress in image analysis and processing III, in Proc. of the
7th Intl. Conf. on Image Analysis and Processing 1993, Singapore, 1994 ,
pp. 547554.
7. A. G. Bharatkumar, K. E. Daigle, M. G. Pandy, Q. Cai, and J. K. Aggarwal,
Lower limb kinematics of human walking with the medial axis transforma-
tion, in Proc. of IEEE Computer Society Workshop on Motion of Non-Rigid
and Articulated Objects, Austin, TX, 1994, pp. 7076.
8. S. A. Niyogiand E. H. Adelson, Analyzingand recognizing walking figures
in XYT, in Proc. IEEE CS Conf. on CVPR, Seattle, WA, 1994, pp. 469474.
9. G. Johansson, Visual motion perception, Sci. Am. 232(6), 1975, 7688.
10. R. F. Rashid, Towards a system for the interpretation of moving light dis-
play, IEEE Trans. PAMI, 2(6), 574581, November 1980.
11. J. A. Webb and J. K. Aggarwal, Visually interpreting the motion of objects
in space, IEEE Comput. August 1981, 4046.12. J. A. Webb and J. K. Aggarwal, Structure from motion of rigid and jointed
objects, in Artif. Intell. 19, 1982, 107130.
13. A. F. Bobick and A. D. Wilson, A state-based technique for the summa-
rization and recognition of gesture, in Proc. of 5th Intl. Conf. on Computer
Vision, 1995, pp. 382388.
14. L. W. Campbell andA. F. Bobick, Recognition of human body motionusing
phasespace constraints,in Proc.of 5thIntl.Conf. onComputerVision, 1995,
pp. 624630.
15. A. Shio and J. Sklansky, Segmentation of people in motion, in Proc. of
IEEE Workshop on Visual Motion, IEEE Computer Society , October 1991,
pp. 325332.
16. B. K. P. Horn and B. G. Schunk, Determining optical flow, Artificial Intel-
ligence, 17, 185204, 1981.17. S. Kurakake and R. Nevatia,Description and tracking of movingarticulated
objects, in 11th Intl. Conf. on Pattern Recognition, Hague, Netherlands,
1992, Vol. 1, pp. 491495.
18. I. A. Kakadiaris, D. Metaxas, and R. Bajcsy, Active part-decomposition,
shape and motion estimation of articulated objects: A physics-based ap-
proach, in Proc. IEEE CS Conf. on CVPR, Seattle, WA, 1994, pp. 980984.
19. I. A. Kakadiaris and D. Metaxas, 3d human body model acquisition from
multiple views, in Proc. of 5th Intl. Conf. on Computer Vision, 1995,
pp. 618623.
20. H. A. Rowley and J. M. Rehg, Analyzing articulated motion using
expectation-maximization, in Proc. of Intl. Conf. on Pattern Recognition,
Puerto Rico, 1997, pp. 935941.
21. A. Jepson and M. Black, Mixture models for optical flow computa
Proc. of Intl. Conf. on Pattern Recognition, New York City, 1993 , pp
761.
22. Z. Chen and H. J. Lee, Knowledge-guided visual perception of 3D
gait from a single image sequence, IEEE Trans. Systems, Man, Cy
22(2), 1992, 336342.
23. E. Huber, 3D real-time gesture recognition using proximity sp
Proc. of Intl. Conf. on Pattern Recognition, Vienna, Austria, Augus
pp. 136141.
24. S. Iwasawa, K. Ebihara, J. Ohya, and S. Morishima, Real-time estiof human body posture from momocular thermal images, in Proc. IE
Conf. on CVPR, Puerto Rico, 1997, pp. 1520.
25. K. Akita, Image sequence analysis of real world human motion, P
Recognit. 17(1), 7383, 1984.
26. F. J. Perales and J. Torres, A system for human motion matching b
synthetic and real image based on a biomechanic graphical model, i
of IEEE Computer Society Workshop on Motion of Non-Rigid and A
lated Objects, Austin, TX, 1994, pp. 8388.
27. M. K. Leung and Y. H. Yang, First sight: A human body outline l
system, IEEE Trans. PAMI17(4), 1995, 359377.
28. J. ORourke andN. I. Badler, Model-based image analysis of human
using constraint propagation, IEEE Trans. PAMI, 2:522536, 1980.
29. D. Hogg, Model-based vision: a program to see a walking person
and Vision Computing, 1(1), 520, 1983.
30. K. Rohr, Towards model-based recognition of human movements in
sequences, CVGIP: Image Understanding 59(1), 94115, 1994.
31. D. Marr and H. K. Nishihara, Representation and recognition of the
organization of three-dimensional shapes, in Proc. R. Soc. London
Vol. B, pp. 269294.
32. S. Wachter and H. H. Nagel, Tracking of persons in monocular
sequences, in Proc. of IEEE Computer Society Workshop on Mo
Non-Rigid and Articulated Objects, Puerto Rico, 1997, pp. 29.
33. J. M. Rehg and T. Kanade, Model-based tracking of self-occluding
lated objects, in Proc. of 5th Intl. Conf. on Computer Vision, Camb
MA, 1995, pp. 612617.34. L. Goncalves, E. D. Bernardo,E. Ursella, andP.Perona, Monoculartr
of the human arm in 3D, in Proc. of 5th Intl. Conf. on Computer
Cambridge, MA, 1995, pp. 764770.
35. I. A. Kakadiaris and D. Metaxas, Model based estimation of 3D
motion with occlusion based on active multi-viewpoint selection, in
IEEE CS Conf. on CVPR, San Francisco, CA, 1996, pp. 8187.
36. D. M. Gavrila and L. S. Davis, Towards 3-D model-based tracki
recognition of human movement: a multi-view approach, in Pro
Workshop on Automatic Face and Gesture Recognition, Zurich, S
land, June 1995.
37. D. M. Gavrila and L. S. Davis, 3-D model-based tracking of hum
action: A multi-view approach, in Proc. IEEE CS Conf. on CVP
Francisco, CA, 1996, pp. 7380.38. H. G. Barrow, Parametric correspondence and chamfer matching: Tw
trchniques for image matching, in Proc. IJCAI, 1977, pp. 659663.
39. E. A. Hunter, P. H. Kelly, and R. C. Jain, Estimation of articulated
using kinematically constrained mixture densities, in Proc. of IEEE
puter Society Workshop on Motion of Non-Rigid and Articulated O
Puerto Rico, 1997, pp. 1017.
40. C. Bregler, Learning and recognizing human dynamics in video sequ
in Proc. IEEE CS Conf. on CVPR, Puerto Rico, 1997, pp. 568574
41. J. K. Aggarwal, L. S. Davis, and W. N. Martin, Correspondence pro
dynamic scene analysis, Proc. IEEE69(5), 1981, 562572.
42. R. Polana and R. Nelson, Low level recognition of human motion (
toget yourman without findinghis bodyparts), in Proc.of IEEE Co
7/29/2019 Q. Cai, Human Motion Analysis a Review
13/13
440 AGGARWAL AND CAI
Society Workshop on Motion of Non-Rigid and Articulated Objects, Austin,
TX, 1994, pp. 7782.
3. Q. Cai, A.Mitiche,and J.K. Aggarwal,Tracking human motionin anindoor
environment, in Proc. of 2nd Intl. Conf. on Image Processing, Washington,
D.C., October 1995, Vol. 1, pp. 215218.
4. J. Segen and S. Pingali, A camera-based system for tracking people in real
time, in Proc. of Intl. Conf. on Pattern Recognition, Vienna, Austria, August
1996, pp. 6367.
5. Y. Okawa and S. Hanatani, Recognition of human body motions by robots,
in Proc. IEEE/RSJ Intl. Conf. Intelligent Robots and Systems, Raleigh, NC,1992, pp. 21392146.
6. M. Rossi and A. Bozzoli, Tracking and counting people, in 1st Intl. Conf.
on Image Processing, Austin, Texas, November 1994, pp. 212216.
7. S. S. Intille and A. F. Bobick, Closed-world tracking, in Proc. Intl. Conf.
Comp. Vis., 1965, pp. 672678.
8. S. S. Intille, J. W. Davis, andA. F. Bobick, Real-time closed-worldtracking,
in Proc. IEEE CS Conf. on CVPR, Puerto Rico, 1997, pp. 697703.
9. C. Wren, A. Azarbayejani, T. Darrel, and A. Pentland, Pfinder: Real-time
tracking of the human body, in Proc. SPIE, Bellingham, WA, 1995.
0. A. Azarbayejani and A. Pentland, Real-time self-calibrating stereo person
tracking using 3D shape estimation from blob features, in Proc. of Intl.
Conf. on Pattern Recognition, Vienna, Austria, August 1996, pp. 627632.
1. B.Gunsel, A. M. Tekalp,and P. J. L. Beek, Object-based video indexing for
virtual studio production, in Proc. IEEE CS Conf. on CVPR, Puerto Rico,
1997, pp. 769774.
2. B. Heisele, U. Kreel, and W. Ritter, Tracking non-rigid moving objects
based on color cluster flow, in Proc. IEEE CS Conf. on CVPR, Puerto Rico,
1997, pp. 257260.
3. J. Yang andA. Waibel, A real-timeface tracker, in Proc. of IEEE Computer
Society Workshop Applications on Computer Vision, Sarasota, FL, 1996,
pp. 142147.
4. P. Fieguth and D. Terzopoulos, Color-based tracking of heads and other
mobile objects at video frame rate, in Proc. IEEE CS Conf. on CVPR,
Puerto Rico, 1997, pp. 2127.
5. D. A. Forsyth and M. M. Fleck, Identifying nude pictures, in Proc. of IEEEComputer Society Workshop Applications on Computer Vision, Sarasota,
FL, 1996, pp. 103108.
6. K. Sato, T. Maeda, H. Kato, and S. Inokuchi, CAD-based object tracking
with distributed monocular camera for security monitoring, in Proc. 2nd
CAD-Based Vision Workshop, Champion, PA, Feburary 1994, p
297.
57. R. Jain and K. Wakimoto, Multiple perspective interactive video,
of Intl. Conf. on Multimedia Computing and Systems, 1995, pp. 20
58. P. H. Kelly, A. Katkere, D. Y. Kuramura, S. Moezzi, S. Chatterj
R. Jain, An architecture for multiple perspective interactive video,
of ACM Conf. on Multimedia, 1995, pp. 201212.
59. Q. Caiand J. K. Aggarwal, Trackinghumanmotion using multiplec
in Proc.of Intl. Conf. on Pattern Recognition, Vienna,Austria, Augu
pp. 6872.
60. Q. Cai and J. K. Aggarwal, Automatic tracking of human motio
door scenes across multiple synchronized video streams, in Intl. C
Computer Vision, Bombay, India, January 1998.
61. Y. Kuno and T. Watanabe, Automatic detection of human for visual
lance system,in Proc. of Intl. Conf.on PatternRecognition,Vienna, A
August 1996, pp. 865869.
62. J. Farmer,M. Casdagli,S. Eubank, and J. Gibson, State-space recons
in the presence of noise, Physics D 51, 1991, 5298.
63. N. H. Goddard, Incremental model-based discriminiation of art
movement from motion features, in Proc. of IEEE Computer Societ
shop on Motion of Non-Rigid and Articulated Objects, Austin, TX
pp. 8995.
64. T. Starner and A. Pentland, Visual recognition of American sign la
using hidden Markov models, in Proc. Intl. Workshop on Automat
and Gesture Recognition, Zurich, Switzerland, June 1995.
65. M. Brand, N. Oliver, and A. Pentland, Coupled hidden Markov mo
complex action recognition, in Proc. IEEE CS Conf. on CVPR,Puer
1997, pp. 994999.
66. A. F. Bobick and J. Davis, Real-time recognition of activity using te
templates, in Proc. of IEEE Computer Society Workshop Applicat
Computer Vision, Sarasota, FL, 1996, pp. 3942.
67. Y. Cui and J. J. Weng, Hand segmentation using learning-based pr
and verification for hand signrecognition,in Proc.IEEE CS Conf. on
Puerto Rico, 1997, pp. 8893.
68. J. E. Boyd andJ. J. Little, Globalversus structured interpretation ofMoving lightdisplays,in Proc.of IEEEComputerSociety Workshop
tion of Non-Rigid and Articulated Objects, Puerto Rico, 1997, pp.
69. A. B. Poritz, Hidden Markov Models: A guided tour., in Proc. IE
Conf. on Acoust. Speech and Signal Proc., 1988, pp. 713.