Gesture Recognition a Survey

Gesture recognition pertains to recognizing meaningful expressions of motion by a human, involving the hands, arms, face, head, and/or body.

It is of utmost importance in designing an intelligent and ef? cient human–computer interface. The applications of gesture recognition are manifold, ranging from sign language through medical rehabilitation to virtual reality.In this paper, we provide a survey on gesture recognition with particular emphasis on hand gestures and facial expressions.

Best services for writing your paper according to Trustpilot

Premium Partner
From $18.00 per page
4,8 / 5
4,80
Writers Experience
4,80
Delivery
4,90
Support
4,70
Price
Recommended Service
From $13.90 per page
4,6 / 5
4,70
Writers Experience
4,70
Delivery
4,60
Support
4,60
Price
From $20.00 per page
4,5 / 5
4,80
Writers Experience
4,50
Delivery
4,40
Support
4,10
Price
* All Partners were chosen among 50+ writing services by our Customer Satisfaction Team

Applications involving hidden Markov models, particle ? ltering and condensation, ? nite-state machines, optical ? ow, skin color, and connectionist models are discussed in detail. Existing challenges and future research possibilities are also highlighted. Index Terms—Face recognition, facial expressions, hand gestures, hidden Markov models (HMMs), soft computing, optical ? ow. I. INTRODUCTION IN THE PRESENT day framework of interactive, intelligent computing, an ef? cient human–computer interaction is assuming utmost importance in our daily lives. Gesture recognition can be termed as an approach in this direction. It is the process by which the gestures made by the user are recognized by the receiver. Gestures are expressive, meaningful body motions involving physical movements of the ? ngers, hands, arms, head, face, or body with the intent of: 1) conveying meaningful information or 2) interacting with the environment.

They constitute one interesting small subspace of possible human motion. A gesture may also be perceived by the environment as a compression technique for the information to be transmitted elsewhere and subsequently reconstructed by the receiver. Gesture recognition has wide-ranging applications [1] such as the following: r developing aids for the hearing impaired; r enabling very young children to interact with computers; r designing techniques for forensic identi? ation; r recognizing sign language; r medically monitoring patients’ emotional states or stress levels; r lie detection; r navigating and/or manipulating in virtual environments; r communicating in video conferencing; r distance learning/tele-teaching assistance; r monitoring automobile drivers’ alertness/drowsiness levels, etc.

Manuscript received June 22, 2004; revised February 28, 2005. This paper was recommended by Editor E. Trucco.

S. Mitra is with the Machine Intelligence Unit, Indian Statistical Institute, Kolkata 700 108, India (e-mail: [email protected] ac. n). T. Acharya is with Avisere Inc. , Tucson, AZ 85287-5706 USA, and also with the Department of Electrical Engineering, Arizona State University, Tempe, AZ 85287-5706 USA (e-mail: [email protected] org).

Digital Object Identi? er 10. 1109/TSMCC. 2007.

893280 Generally, there exist many-to-one mappings from concepts to gestures and vice versa. Hence, gestures are ambiguous and incompletely speci? ed. For example, to indicate the concept “stop,” one can use gestures such as a raised hand with palm facing forward, or, an exaggerated waving of both hands over the head.Similar to speech and handwriting, gestures vary between individuals, and even for the same individual between different instances. There have been varied approaches to handle gesture recognition [2], ranging from mathematical models based on hidden Markov chains [3] to tools or approaches based on soft computing [4].

In addition to the theoretical aspects, any practical implementation of gesture recognition typically requires the use of different imaging and tracking devices or gadgets.These include instrumented gloves, body suits, and markerbased optical tracking. Traditional 2-D keyboard-, pen-, and mouse-oriented graphical user interfaces are often not suitable for working in virtual environments. Rather, devices that sense body (e. g.

, hand, head) position and orientation, direction of gaze, speech and sound, facial expression, galvanic skin response, and other aspects of human behavior or state can be used to model communication between a human and the environment.Gestures can be static (the user assumes a certain pose or con? guration) or dynamic (with prestroke, stroke, and poststroke phases). Some gestures also have both static and dynamic elements, as in sign languages. Again, the automatic recognition of natural continuous gestures requires their temporal segmentation. Often one needs to specify the start and end points of a gesture in terms of the frames of movement, both in time and in space. Sometimes a gesture is also affected by the context of preceding as well as following gestures.Moreover, gestures are often language- and culture-speci? c.

They can broadly be of the following types: 1) hand and arm gestures: recognition of hand poses, sign languages, and entertainment applications (allowing children to play and interact in virtual environments); 2) head and face gestures: some examples are: a) nodding or shaking of head; b) direction of eye gaze; c) raising the eyebrows; d) opening the mouth to speak; e) winking, f) ? aring the nostrils; and g) looks of surprise, happiness, disgust, fear, anger, sadness, contempt, etc. 3) body gestures: involvement of full body motion, as in: a) tracking movements of two people interacting outdoors; b) analyzing movements of a dancer for generating matching music and graphics; and c) recognizing human gaits for medical rehabilitation and athletic training. Typically, the meaning of a gesture can be dependent on the following: r spatial information: where it occurs; r pathic information: the path it takes; 1094-6977/$25.

00 © 2007 IEEE 312 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 3, MAY 2007 symbolic information: the sign it makes; r affective information: its emotional quality. Facial expressions involve extracting sensitive features (related to emotional state) from facial landmarks such as regions surrounding the mouth, nose, and eyes of a normalized image. Often dynamic image frames of these regions are tracked to generate suitable features.

The location, intensity, and dynamics of the facial actions are important for recognizing an expression. Moreover, the intensity measurement of spontaneous facial expressions is often more dif? cult than that of posed facial expressions.More subtle cues such as hand tension, overall muscle tension, locations of self-contact, and pupil dilation are sometimes used. In order to determine all these aspects, the human body position, con? guration (angles and rotations), and movement (velocities) need to be sensed. This can be done either by using sensing devices attached to the user. Those may be magnetic ? eld trackers, instrumented (data) gloves, and body suits, or by using cameras and computer vision techniques. Each sensing technology varies along several dimensions, including accuracy, resolution, latency, range of motion, user comfort, and cost.Glove-based gestural interfaces typically require the user to wear a cumbersome device and carry a load of cables connecting the device to a computer.

This hinders the ease and naturalness of the user’s interaction with the computer. Vision-based techniques, while overcoming this, need to contend with other problems related to occlusion of parts of the user’s body. While tracking devices can detect fast and subtle movements of the ? ngers when the user’s hand is moving, a vision-based system will at best get a general sense of the type of ? nger motion.Again, vision-based devices can handle properties such as texture and color for analyzing a gesture, while tracking devices cannot.

Vision-based techniques can also vary among themselves in: 1) the number of cameras used; 2) their speed and latency; 3) the structure of environment (restrictions such as lighting or speed of movement); 4) any user requirements (whether user must wear anything special); 5) the low-level features used (edges, regions, silhouettes, moments, histograms); 6) whether 2-D or 3-D representation is used; and 7) whether time is represented.There is, however, an inherent loss in information whenever a 3-D image is projected to a 2-D plane. Again, elaborate 3-D models involve prohibitive highdimensional parameter spaces. A tracker also needs to handle changing shapes and sizes of the gesture-generating object (that varies between individuals), other moving objects in the background, and noise.

Good review on human motion analysis is available in literature [5], [6]. In this paper, we provide a survey on different aspects of gesture recognition. Section II outlines various tools often used for gesture recognition.Section III is devoted to hand and arm gestures, with particular emphasis on hidden Markov models (HMMs), particle ? ltering and condensation, ? nite-state machine (FSM), and neural network. This is followed by facial gesture recognition in Section IV including a coverage on approaches employing HMMs, principal component analysis (PCA), contour models, feature extraction, Gabor ? ltering, optical ? ow, skin color, and connectionist models. Finally, Section V indicates a few of the existing challenges and future research possibilities. II.

TOOLS FOR GESTURE RECOGNITION Gesture recognition is an ideal example of multidisciplinary research. There are different tools for gesture recognition, based on the approaches ranging from statistical modeling, computer vision and pattern recognition, image processing, connectionist systems, etc. Most of the problems have been addressed based on statistical modeling, such as PCA, HMMs [3], [7], [8], Kalman ? ltering [9], more advanced particle ? ltering [10], [11] and condensation algorithms [12]–[14]. FSM has been effectively employed in modeling human gestures [15]–[18].Computer vision and pattern recognition techniques [19], involving feature extraction, object detection, clustering, and classi? cation, have been successfully used for many gesture recognition systems. Image-processing techniques [20] such as analysis and detection of shape, texture, color, motion, optical ? ow, image enhancement, segmentation, and contour modeling [21], have also been found to be effective. Connectionist approaches [22], involving multilayer perceptron (MLP), timedelay neural network (TDNN), and radial basis function network (RBFN), have been utilized in gesture recognition as well.While static gesture (pose) recognition can typically be accomplished by template matching, standard pattern recognition, and neural networks, the dynamic gesture recognition problem involves the use of techniques such as time-compressing templates, dynamic time warping, HMMs, and TDNN.

In the rest of this section, we discuss the principles and background of some of these popular tools used in gesture recognition. A. HMM A time-domain process demonstrates a Markov property if the conditional probability density of the current event, given all present and past events, depends only on the jth most recent event.If the current event depends solely on the most recent past event, then the process is termed a ? rst order Markov process.

This is a useful assumption to make, when considering the positions and orientations of the hands of a gesturer through time. The HMM [3], [7] is a double stochastic process governed by: 1) an underlying Markov chain with a ? nite number of states and 2) a set of random functions, each associated with one state. In discrete time instants, the process is in one of the states and generates an observation symbol according to the random function corresponding to the current state.Each transition between the states has a pair of probabilities, de? ned as follows: 1) transition probability, which provides the probability for undergoing the transition; 2) output probability, which de? nes the conditional probability of emitting an output symbol from a ? nite alphabet when given a state. The HMM is rich in mathematical structures and has been found to ef? ciently model spatio–temporal information in a natural way. The model is termed “hidden” because all that can be MITRA AND ACHARYA: GESTURE RECOGNITION: A SURVEY 313 Fig.

. Five-state left-to-right HMM for gesture recognition. Given an observation sequence, the following are the key issues in HMM use: 1) evaluation: determining the probability that the observed sequence was generated by the model (Forward– Backward algorithm); 2) training or estimation: adjusting the model to maximize the probabilities (Baum–Welch algorithm); 3) decoding: recovering the state sequence (Viterbi algorithm).

B. Particle Filtering and Condensation Algorithm seen is only a sequence of observations. It also involves elegant and ef? ient algorithms, such as Baum–Welch and Viterbi [23], for evaluation, learning, and decoding. An HMM is expressed as ? = (A, B, ? ) and is described as follows: r a set of observation strings O = {O1 , . . .

, OT }, where t = 1, . . . , T ; r a set of N states {s1 , .

. . , sN }; r a set of k discrete observation symbols {v1 , .

. . , vk }; r a state-transition matrix A = {aij }, where aij is the transition probability from state si at time t to state sj at time t+1 A={aij }=Prob(sj at t + 1|si at t), for 1 ? i, j ? N. (1) r an observation symbol probability matrix B = {bjk }, here bjk is the probability of generating symbol vk from state sj ; r an initial probability distribution for the states ? ={? j }, j =1, 2, . . . , N, where ? j = Prob(sj at t = 1). The generalized topology of an HMM is a fully connected structure, known as an ergodic model, where any state can be reached from any other state.

When employed in dynamic gesture recognition, the state index transits only from left to right with time, as depicted in Fig. 1. The start state s1 and ? nal state sN , for N = 5, are indicated on the ? gure. Here, the statetransition coef? ients aij = 0 if j < i, and N aij = 1. The j=1 Viterbi algorithm is used for evaluating a set of HMMs and decoding by considering only the maximum path at each time step instead of all paths. The global structure of the HMM is constructed by parallel connections of each HMM (? 1 , ? 2 , . . .

, ? M ), whereby insertion (or deletion) of a new (or existing) HMM is easily accomplished. Here, ? corresponds to a constructed HMM model for each gesture, where M is the total number of gestures being recognized. HMMs have been applied to hand and face recognition.Usually, a 2-D projection is taken from the 3-D model of the hand or face, and a set of input features are extracted experimentally. The spatial component of the dynamic gesture is typically neglected, while the temporal component (having a start state, end state, and a set of observation sequences) is mapped through an HMM classi? er with appropriate boundary conditions. A set of data are employed to train the classi? er, and the test data are used for prediction veri? cation. Particle-? ltering-based tracking and its applications in gesture recognition systems became popular very recently [10]–[14].Particle ? lters have been very effective in estimating the state of dynamic systems from sensor information.

The key idea is to represent probability densities by set of samples. As a result, it has the ability to represent a wide range of probability densities, allowing real-time estimation of nonlinear, non-Gaussian dynamic systems. This technique was originally developed to effectively track objects in clutter [12], [13]. The state of a tracked object at time t is described by a vector Xt , where the vector Yt represents all the samples of observations {y1 , y2 , · · · , yt }.The posterior density P (Xt |Yt ) and the observation density P (Yt |Xt ) are often non-Gaussian.

Basically, the particle ? lters are a sample-based variant of Bayes ? lters. The key idea is to approximate the probability density distribution by a weighted sample set St = (i) (i) (i) { xt , wt | i = 1, · · · , Np }. Here, each sample xt represents (i) a hypothetical state of the object, and wt represents the corre(i) sponding discrete sampling probability of the sample xt such that Np wt = 1. i=1 (i) (2) The particle ? ltering in its basic form actually realizes the recursive Bayes ? ter according to a sampling procedure, often called sequential importance sampling with resampling (SISR) [14]. The iterative evolution of the sample set is described by propagating each sample according to a system model. Each sample element in the set is weighted in terms of the observations, and Np samples are drawn with replacement by choosing a particu(i) (i) lar sample with posterior probability wt = P (yt | Xt = xt ).

In each step of iteration, the mean state (sample) of an object is estimated as Np E(St ) = i=1 wt xt . (i) (i)Since it models uncertainty (as posterior probability density), particle ? ltering provides a robust tracking framework suitable for gesture recognition systems. Based on the above principle, the condensation algorithm (conditional density propagation over time) was originally proposed to deal with the problem of tracking rapid motion in clutter [13]. Rather than attempting to ? t a speci? c equation to the observed sensory data, it uses the Np weighted samples to approximate the curve described by the observed data. When applied to tracking, each sample 314IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL.

37, NO. 3, MAY 2007 represents the state of the object being tracked, e. g. , its velocity and location. Given such a randomly sampled state St at time t, a prediction of a new state St+1 at time t + 1 is made using a predictive model. An extension of the condensation algorithm has been proposed [24] to automatically switch between various prediction motion models. Because of the usage of multiple such models to predict different types of motion of the objects, this signi? antly improves the performance of the tracker.

It also enables the extraction of the most likely particular model that correctly represents the motion observed at a given time t. This is very relevant and useful for gesture recognition. If each competing model represents a single gesture, then the most likely model predicts which gesture is being observed. C. FSM Approach In the FSM approach, a gesture can be modeled as an ordered sequence of states in a spatio–temporal con? guration space [15]–[18].

The number of states in the FSM may vary between applications.The gesture is recognized as a prototype trajectory from an unsegmented, continuous stream of sensor data constituting an ensemble of trajectories. The trajectories of the gestures are represented as a set of points (e. g. , sampled positions of the head, hand, and eyes) in a 2-D space.

Usually, the training of the model is done off-line, using many possible examples of each gesture as training data, and the parameters (criteria or characteristics) of each state in the FSM are derived. The recognition of gestures can be performed online using the trained FSM.When input data (feature vectors such as trajectories) are supplied to the gesture recognizer, the latter decides whether to stay at the current state of the FSM or jump to the next state based on the parameters of the input data. If it reaches a ? nal state, we say that a gesture has been recognized. The state-based representation can be extended to accommodate multiple models for the representation of different gestures, or even different phases of the same gesture. Membership in a state is determined by how well the state models can represent the current observation.

If more than one model (gesture recognizer) reach their ? al states at the same time, we can apply a winning criteria to choose the most probable gesture. The concept of motion energy has been used [17] to extract the temporal signature of hand motion from a limited set of dynamic hand gestures. This is subsequently analyzed and interpreted by a deterministic FSM. The relative change of direction of motion, rather than the relative motion such as quickly or slowly, is considered to be more important for determining the temporal signature. Adaptation with cross-cultural gestures can be achieved by rede? ning the FSM according to the relevant rules of the society.Inclusion of new gestures is achieved by the construction of additional FSMs portraying the corresponding motion pro? le. D.

Soft Computing and Conectionist Approach Soft computing is a consortium of methodologies that works synergistically and provides ? exible information processing capability for handling real-life ambiguous situations [25]. Its aim is to exploit the tolerance for imprecision, uncertainty, approximate reasoning, and partial truth in order to achieve tractability, robustness, and low-cost solutions. Sensor outputs are often associated with an inherent uncertainty.

Relevant, sensor-independent, invariant features are extracted from these outputs, followed by gesture classi? cation. The recognition system may be designed to be fully trained when in use, or may adapt dynamically to the current user. Soft computing tools [26], such as fuzzy sets, arti? cial neural networks (ANNs), genetic algorithms (GAs), and rough sets, hold promise in effectively handling these issues. Management of uncertainty at different levels can be made by fuzzy sets and rough sets, while dimensionality reduction is taken care of by self-organizing maps.Fuzzy sets can also be used to model simultaneous partial membership to multiple gesture classes. The adaptive nature of ANNs enable connectionist approaches to incorporate learning in data-rich environment. This characteristic, coupled with robustness, is useful in developing recognition systems. The various connectionist models in literature [27]–[30], for hand gesture and facial expression recognition, include MLP, TDNN, and RBFN.

III. HAND AND ARM GESTURES Human gestures typically constitute a space of motion expressed by the body, face, and/or hands.Of these, hand gestures are often the most expressive and the most frequently used.

This involves: 1) a posture: static ? nger con? guration without hand movement and 2) a gesture: dynamic hand movement, with or without ? nger motion. Gestures may be categorized as given in the following list, such that as we proceed downward this list, their association with speech declines, language properties increase, spontaneity decreases, and social regulation increases: r gesticulation: spontaneous movement of hands and arms, accompanying speech. These spontaneous movements constitute around 90% of human gestures.People gesticulate when they are on telephone, and even blind people regularly gesture when speaking to one another; r languagelike gestures: gesticulation integrated into a spoken utterance, replacing a particular spoken word or phrase; r pantomimes: gestures depicting objects or actions, with or without accompanying speech; r emblems: familiar signs such as “V for victory,” or other culture-speci? c “rude” gestures; r ign languages: well-de? ned linguistic systems. These carry the most semantic meaning and are more systematic, thereby being easier to model in a virtual environment.Hand gesture recognition consists of gesture spotting that implies determining the start and end points of a meaningful gesture pattern from a continuous stream of input signals and, subsequently, segmenting the relevant gesture. This task is very dif? cult due to: 1) the segmentation ambiguity and 2) the spatio– temporal variability involved. As the hand motion switches from one gesture to another, there occur intermediate movements as well.

These transition motions are also likely to be segmented and matched with reference patterns, and need to be eliminated MITRA AND ACHARYA: GESTURE RECOGNITION: A SURVEY 15 by the model. Moreover, the same gesture may dynamically vary in shape and duration even for the same gesturer. Orientation histograms have also been used as a feature vector for fast, simple hand gesture classi? cation and interpolation [31]. It is based on computing the Euclidean distance from the prototypes in terms of these features.

The histogram of orientations, representing an orientation in terms of its angle, provides for translational invariance. A. HMMs for Hand Gesture Recognition HMM is a rich tool used for hand gesture recognition in diverse application domains.Probably, the ? rst publication addressing the problem of hand gesture recognition is the celebrated paper by Yamato et al. [7]. In this approach, a discrete HMM and a sequence of vector-quantized (VQ)-labels have been used to recognize six classes of tennis strokes. Before applying the HMM, the image sequence goes through several preprocessing steps such as low-pass ? ltering to reduce the noise, background subtraction to extract the moving objects, and binarization of the moving objects in order to generate blobs. The blobs roughly represent the poses of the human.

The features are the amounts of object (black) pixels. These features are vector quantized, such that the image sequence becomes a sequence of VQ-labels, which are then processed by a discrete HMM. Subsequently, several other applications of hand gesture recognition have been developed based on HMMs. In the following sections, we elaborate some of these, as available in the literature. 1) Sign Language Recognition: Sign language is a visual language. Usually, sign language consists of three major components: 1) ? nger-spelling; 2) word-level sign vocabulary; and 3) nonmanual features [32].The ? nger-spelling is used to spell words letter by letter.

The word-level sign vocabulary is used for majority of communication, while the nonmanual features consist of facial expression, position of tongue, mouth, and body. A real-time HMM-based system has been designed [33], [34] for recognizing sentence-level American Sign Language, without explicitly modeling the ? ngers. Since each gesture in a sign language has an already assigned meaning, strong rules of context and grammar may be applied to make the recognition tractable.Typically, the gestures represent whole words. A ?vestate HMM is used to recognize data strings, and is combined with statistical grammar to incorporate context during training and recognition. The Viterbi algorithm is used both with and without a strong grammar, based on the known forms of the sentences. However, once the recognizer starts, the subject must conduct only sign languages. This is because the model cannot distinguish unde? ned hand motions.

The subject wears distinctly colored gloves on both hands, and sits in a chair in front of the camera to aid hand tracking.An eight-component feature vector consists of the hand’s 2-D coordinates, axis angle of least inertia, and the eccentricity of the bounding ellipse. Considering that all human hands have approximately the same hue and saturation but vary in their brightness, in another approach, the hands are tracked based on skin tone.

The leftmost and rightmost hands are assigned to be “left” and “right,” respectively. The HMM-based approach, described above, requires extensive training sets for modeling. It was found to be effective for practically just around 50 words and required heavily constrained arti? ial grammar on the structure of the sentences [33], [35].

Bowden et al. [32], [35] made signi? cant progress in sign language interpretation by structuring the classi? cation model around a linguistic de? nition of signed words, rather than an HMM. In this vision-based approach, the “visemes” of sign (aka phase representation of constituent motions) are de? ned in a manner similar to that in the sign linguistic dictionary. Viseme is a visual equivalent of a phoneme. This enables signs to be learnt reliably from just a handful of training examples.The system is designed based on a novel two-stage classi? cation.

In the ? rst stage, the raw image sequences are segmented in order to extract shapes and trajectories of the hands in the binary image sequence. These features are then converted into a viseme representation: 1) position of the hands relative to each other (HA); 2) position of hands relative to key body locations (TAB); 3) relative movement of the hands (SIG); and 4) the shape of the hands (DEZ). In the second stage of classi? cation, each sign is modeled as a ? st order Markov chain in which each state in the chain represents a particular set of feature vectors generated from the classi? cation in stage one.

The Markov chain encodes temporal transitions of the hands. In order to allow minor variations over the sign instances and optimal featureto-symbol mapping, independent component analysis (ICA) is employed to separate the correlated features from uncorrelated noise. Once the ICA transformation matrix has been learnt, a lookup table (LUT) is generated from the training data to map the ICA transformed features to symbols for use in Markov chain.The salient feature of the above work [32], [35] is that the Markov chains can be built from as little as a single training example, or alternatively, a hand-coded description of the sign, due to generalization of the features. It still yields results comparable to the state-of-the-art. 2) Graphic Editor Control: Another HMM-based model [36] uses hand localization, hand tracking, and gesture spotting at preprocessing for hand gesture recognition.

Hand candidate regions are located on the basis of skin color and motion.The centroids of the moving hand regions are connected to produce a hand trajectory, which is then divided into real and meaningless segments (categories). Input feature codes are extracted in terms of combined weighted location, angle, and velocity. This is followed by c-means clustering [37] to generate the HMM codebook.

Left-to-right HMM with ten states is used for recognizing hand gestures to control a graphic editor. The gestures modeled include 12 graphic elements (circle, triangle, rectangle, arc, horizontal line, and vertical line) and 36 alphanumeric characters (ten Arabic numerals and 26 alphabets).A charge-coupled device (CCD) camera placed in front of a monitor gives a sequence of gesture images from an imagecapture board. The I and Q components of the YIQ color system are used here to extract hand areas from input images. A priori knowledge about the hand location of a previous video image, the usual face location, and the size of the hand region are used 316 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 37, NO.

3, MAY 2007 to distinguish the hand region from multiple candidate regions.The basic hand location algorithm is outlined as follows: 1) color system conversion from RGB to YIQ; 2) estimation of similarity measures between model and input regions; 3) thresholding similarity measures; 4) noise removal and dilation; 5) detection of hand candidate regions; 6) selection of hand region. Garbage movements that come before and after a pure gesture are removed by using a spotting rule, whereby the user intentionally stops for a while (2–3 s) at the start and end of the gesture.

An eight-connectivity counterclockwise directional chain code is used to convert the orientation angles into feature codes.The velocity component takes care of the fact that while a simple “circle” gesture may have an almost nonvarying speed, a complex “q” or “w” gesture generation can involve varying speeds of movement. 3) Robot Control: A combination of static shape recognition, Kalman-? lter-based hand tracking, and an HMM-based temporal characterization scheme is developed [38] for reliable recognition of single-handed dynamic hand gestures.

Here, the start and end of gesture sequences are automatically detected. Hand poses can undergo motion and discrete changes in between the gesture, thereby enabling one to deal with a larger set of gestures.However, any continuous deformation of hand shapes is not allowed. The system is robust to background clutter, and uses skin color for static shape recognition and tracking. A real-time implementation is developed for robot control. The user is expected to bring the hand to a designated region for initiating a gestural action to be grabbed by a camera. The hand should be properly illuminated, and the user should avoid sudden jerky movements. When the user moves the hand away from the designated region, it signals the end of the gesture and the beginning of the recognition process.

The grabber and tracker are operated as synchronized threads. The ? ve hand gestures and the corresponding instructions modeled for the robot are: 1) closed to open forward: move forward; 2) closed to open right: move forward then right; 3) closed to open left: move forward then left; 4) open to closed right: move backward then right; and 5) open to closed left: move backward then left. Static hand shapes are described by their contours, speci? ed through mouse clicks on the boundary of the image, and subsequently ? tting a B-spline curve. Translated, scaled, and rotated versions of these shapes are also added to the prior.For a test shape, a matching is made with these stored priors. The closest contour match is chosen for tracker initialization. A gesture is considered as a sequence of epochs, where each epoch is characterized by a motion of distinct hand shapes.

Kalman ? lter is used for hand tracking, to obtain motion descriptors for the HMM. The moving hand is approximated as a planar rigid shape, under the assumption that the ? ngers are not being ? exed, and the perspective effects are not signi? cant. The left-to-right HMM, with four states and an out degree of three, proceeds by doing the following: extracting symbolic descriptors of the gesture at regular r training HMMs by the sequence of symbolic descriptors r ? nding the model, which gives maximum probability of occurrence of the observation sequence generated by the test gesture. The gesture recognition algorithm is outlined as follows. 1) Detect hand for boot-strapping the tracker. 2) Recognize the starting hand shape, and initialize tracker with its template. 3) While hand is in view repeat a) Track the hand and output encoded motion information until shape change is detected. ) Recognize the new shape and initialize the tracker with template of the recognized shape.

4) Using HMM, ? nd the gesture, which gives the maximum probability of occurrence of observation sequence composed of shape templates and motion information. B. Condensation Algorithm The condensation algorithm was developed based on the principle of particle ? ltering. It was originally applied effectively in tracking rapid motion of objects in clutter [13]. A mixed-state condensation algorithm has been extended to recognize a greater number of gestures based on their temporal trajectories [39].

Here, one of the gesture models involves an augmented of? ce white-board with which a user can make simple hand gestures to grab regions of the board, print them, save them, etc. In this approach, the authors allow compound models that are very like HMMs, with each state in the HMM being one of the de? ned trajectory models. The other part deals with human facial expressions, using the estimated parameters of a learned model of mouth motion [39]. C. FSMs for Hand Gesture Recognition As discussed in Section II-C, a gesture can be modeled as an ordered sequence of states in a spatio–temporal con? uration space in the FSM approach. This has been used to recognize hand gestures [15], [17], [18]. A method to recognize human-hand gestures using a FSMmodel-based approach has been used in [15].

The state machine is used to model four qualitatively distinct phases of a generic gesture—static start position (static at least for three frames), smooth motion of the hand and ? ngers until the end of the gesture, static end position for at least three frames, and smooth motion of the hand back to the start position.The hand gestures are represented as a list of gesture vectors and are matched with the stored gesture vector models based on vector displacements. Another state-based approach to gesture learning and recognition has been presented in [18]. Here, each gesture is de? ned to be an ordered sequence of states, using spatial clustering and temporal alignment. The spatial information is ? rst learned from a number of training images of the gestures.

This information is used to build FSMs corresponding to each gesture. The FSM is corresponding to each gesture; intervals from the tracker and hand shape classi? r; MITRA AND ACHARYA: GESTURE RECOGNITION: A SURVEY 317 then used to recognize gestures from an unknown input image sequence. In [17], a human performs a gesture in the ? eld of view of a strategically placed camera, attached to a standard frame grabbing and digitizing hardware, against a simple stationary background. The gesture is started from any arbitrary spatio– temporal position and is a continuous stream of captured motion.

The grabbing system captures the data with a spatial resolution of 256 ? 256 pixels, at a rate of 15 frames/s.The gesture data are temporally segmented into subsequences involving uniform dynamics along a single direction (such as left, right, etc. ). The system is highly recon? gurable, and no training concept is involved.

The FSM has ? ve states, viz. , start (S), up (U), down (D), left (L), and right (R). The length of the signature pattern is independent of the duration over which the gesture is performed, but depends on the number of changes in the dominant direction of motion. Self-loops are essential to accommodate the idleness of hand movement while changing the direction of hand waving.Symbolic commands such as come closer, go far, move left, move right, and emergency stop are recognized.

For example, come closer is modeled by repeatedly sweeping one hand toward the body and then slowly away (say, by S-D-U-U-D-U-D-D-UD). Again, move right may be represented by moving the hand continuously toward the right direction and then the left (say, by S-L-R-R-L-L-R-L-R-L). A lexicon is constructed to model the motion pro? le of the hand for each gesture. This knowledge is utilized for signature representation in the form of simple production rules, followed by their interpretation.These can be used as input to a robot programming language for generating machine-level instructions, in order to mimic the intended operation of the corresponding gesture.

D. Connectionist Approach to Hand Gesture Recognition TDNN has been applied in [27] to recognize gestures related to American Sign Language (ASL). The gesture recognition system in this approach has been divided into two distinct steps. In the ? rst step, multiscale motion segmentation is applied to track movements of objects between the frames. Regions between two consecutive frames are matched to obtain two-view correspondences. Then, af? e transformations are computed to de? ne pixel matches and recognize the motion regions or trajectories.

In the second step, a TDNN is used to match the trajectory of movements to a given gesture model. To recognize the sign language, only those object areas of motion where skin color is detected are determined ? rst. Then, the detected regions of motion are merged until the shape of the merged region is either an ellipse or a rectangle. This is because sign languages are typically described by the relation between head (a large elliptical shape), palm (small elliptical shape), and (or) closed hand (a rectangular shape).The TDNN with two hidden layers is employed to classify the motion of various regions over time as a particular gesture (sign) in the sign language (ASL). The experiment was done with a database of 40 complex hand gestures of ASL and each sign with around 38 instances on the average. Each video consists of an ASL sign which lasts for about 3 to 5 s at 30 frames/s with 160 ? 120 pixels image resolution of each frame.

The recognition rate achieved on the training and testing sets for gesture recognition are a high 99. 02% and 96. 21%, respectively. IV.FACE AND HEAD GESTURES Face is a unique feature of a human being. Humans can detect and identify faces in a scene with little or no effort. Their robustness is tremendous, considering the large changes inherent in the visual stimulus due to: 1) viewing conditions (such as variation in luminance); 2) facial expression; 3) aging; 4) gender; 5) occlusion; or 6) distractions such as glasses, hair style or other disguise.

Human faces are nonrigid objects with a high degree of variability in size, shape, color, and texture. The goal of face detection is to ef? iently identify and locate human faces regardless of their positions, scales, orientations, poses, and illumination. Any automated system for face and facial gesture recognition will have immense potential in criminal identi? cation, surveillance, missing children retrieval, of? ce security, credit card veri? cation, video document retrieval, telecommunication, high-de? nition television (HDTV), medicine, human–computer interfaces, multimedia facial queries, and low-bandwidth transmission of facial data [40]–[42]. While frontal recognition is the classical approach, pro? e recognition schemes are practical for a fast, coarse presearch of large databases in order to reduce computational complexity for a subsequent sophisticated algorithm.

One needs to concentrate on the extraction of appropriate image attributes with useful query functionality, retrieval methods on similarity-based (instead of exact) match, query-by-image example (content-based image retrieval (CBIR)), query re? nement, and high-dimensional database indexing. There are two major approaches to automated face recognition [43]. r Analytic: here ? xible mathematical models are developed to incorporate face deformation and illumination changes. Discrete local (geometrical) features, such as irises and nostrils, are extracted for retrieving and identifying faces. The position of these features, with respect to one another, determine the overall location of the face. Standard statistical pattern recognition techniques such as HMMs [8], may be applied on these measurements. Other approaches include active contour models (Snakes) [21], wavelets [44], and knowledge- or rule-based techniques such as facial action coding system (FACS) [28], [45]. Holistic: this involves gray-level template matching using global recognition.

Here a feature vector is used to represent the entire face template. This approach includes ANNs [28]–[30], [46], [49], linear discriminants, PCA, singular value decomposition (SVD) using eigenfaces [48], [49], and optical ? ow [30], [50]–[52]. Research indicates that a fusion of both approaches often produces better (more stable) results, given an observation sequence, as compared to either approach alone [53]. Facial features can, again, be of two types [54].They are: 1) permanent 318 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 37, NO. 3, MAY 2007 facial features (such as eyes, eyebrows, lips, cheeks, tissue texture, facial hair, and permanent furrows) that are always present in the face, but may be deformed due to facial expressions and 2) transient facial features (such as wrinkles, furrows, and bulges) that occur in the forefront and the regions surrounding the mouth and eyes.

Variation between two face images can be twofold, viz. 1) interpersonal—difference in appearance due to different identity and 2) intrapersonal—change in appearance of the same person due to different facial expressions or lighting. The ? rst category corresponds to face recognition, while the second one pertains to facial expression recognition.

Facial expressions can again be classi? ed in terms of the following: 1) facial actions that cause an expression; 2) nonprototypic expressions such as “raised brows”; or 3) prototypic expressions such as in different emotions. It is interesting to note that the upper face features play a more important role in expression classi? ation as compared to the lower face features. The issue of CBIR has been addressed on the facial recognition technology (FERET)1 facial database [47]. The FERET database consists of frontal shots as well as right and left pro? les of human faces. A dif? culty posed by this database is that the images were often taken at different times, locations, and under different imaging conditions. Moreover, the “gallery” subset (training set) and the “probe” subset (test set) were captured about a week apart and exhibit differences in clothing, hair, and lighting.A. HMMs in Face Recognition HMMs have been applied for interpersonal face recognition [8].

A sliding window is applied from the top to bottom of a 2-D image, while extracting the brightness value of the window as the 1-D feature vectors for the HMM. Successive windows are overlapped to avoid cutting of signi? cant facial features and to provide contextual information for the sequence of feature vectors. The face is modeled as a linear left-to-right HMM, with ? ve states corresponding to the important regions such as forehead, eyes, nose, mouth, and chin.The algorithm is outlined as follows. r Build a code book from the feature vectors of a set of images. r Quantize feature vectors from this code book. r Train HMM for every person in the database. r Recognize a test image, preprocessed using the ? rst two steps.

B. PCA The eigenfaces method [48] is based on the statistical representation of the face space. It ? nds the principal components (Karhunen–Loeve expansion) of the facial image distribution, or, the eigenvectors of the covariance matrix of the set of face images.These eigenvectors, representing a set of macrofeatures (that are generated a posteriori on a statistical basis) characterizing the face, constitute the eigenfaces.

Fisher’s linear discrim1 http://www. itl. nist.

gov/iad/humanid/feret/ inant is used to develop a set of feature vectors, constituting Fisherfaces, where the interpersonal variations are emphasized as compared to the intrapersonal variations. Essa and Pentland [49] employed eigenfaces approximated by PCA, to locate faces in an arbitrary scene.The distance of an observed image from the face space is calculated using the projection coef? cients and the signal energy. Facial motion is extracted using holistic dense optical ? ow, coupled with 3-D motion and muscle-based face models. Faces are detected from an image sequence by performing a spatio–temporal ? ltering, followed by thresholding, in order to analyze “motion blobs” that represent human heads. A control-theoretic method is used to extract the spatio–temporal motion-energy representation of facial motion for an observed expression.The Euclidean distance is computed from the “ideal” motion-energy templates for the six different expressions, involving two facial actions (viz.

, smile and raised eyebrows) and four emotional expressions (viz. , surprise, disgust, sadness, and anger). A correct recognition rate of 98% is reported with 52 frontal-view image sequences of eight people. C. FACS FACS [45] was designed to help human observers detect independent subtle changes in facial appearance caused by contractions of the facial muscles.

FACS provides linguistic rules describing all possible, visually detectable facial changes in terms of 44 action units (AUs). Thirty AUs are anatomically related to the contractions of speci? c facial muscles, viz, 12 for the upper face and 18 for the lower face. The AUs can occur either singly or in combination. For example, in FACS, an “inner brow raiser” corresponds to AU1 while a “jaw drop” refers to AU26. Using these rules, a trained human FACS coder decomposes a facial expression into speci? c AUs describing it.

However, here it is not possible to classify an xpression into multiple emotion categories; for example, raised eyebrows and smiling mouth being expressed as a blend of surprise and happiness. D. Contour Models Facial features such as lips, eyebrows, and nose are often extracted using active contour models (Snakes) [21].

This is an energy-minimizing spline guided by external constraint forces and affected by image forces that pull it toward features such as lines and edges. The local minima of the energy function correspond to the set of solutions. Addition of energy terms push the active model toward the desired solution when placed near it.

Snakes are initialized by placing them on the facial features that are to be tracked. They lock onto nearby edges by localizing them accurately. and are able to deform and accurately track the features. E. Facial Feature Extraction for Gesture Recognition A combined approach is used for facial feature extraction and determination of gaze direction [55], as applied to facial expression recognition. The features under consideration here MITRA AND ACHARYA: GESTURE RECOGNITION: A SURVEY 319 are the eyebrows, eyes, nostrils, mouth, cheeks, and chin.The model employs the following: 1) an improved variation of adaptive Hough transform [56] for geometrical shape parameterization, involving curve detection based on ellipse containing the main (oval) connected component of the image (related to cheeks and chin detection); 2) minima analysis of feature candidates corresponding to the low-intensity regions of the face (extracting eyes, nostrils, and mouth); 3) template matching for inner facial feature localization, using an appropriate binary mask on an area restricted by the eyes (extracting upper eyebrow edges that may not otherwise be uniformly described by a geometric curve); 4) dynamic deformation of active contours for inner face contour detection; 5) projective geometry properties for accurate pose determination, along with analysis of face symmetry properties for determination of gaze direction. The skinlike regions in the image are detected from the huesaturation-value (HSV) color space representation.

Several classes of perceptual cue to emotional states are displayed on the face.These include the following: r the relative displacement of features [44] such as opening the mouth; r quasi-textural changes in the skin surface such as furrowing the brow [44]; r changes in skin hue such as blushing; r temporal aspect of these signals, involving motion [30], [51]. While motion extraction approaches directly focus on facial changes occurring due to facial expressions, the deformationbased methods rely on neutral face images (or models) to extract facial features that are relevant to facial actions and are not caused by (say) intransient wrinkles due to old age. 1) Uni? ed Parameterized Facial Appearance Model for Recognition: Lanitis et al. 57] have designed a compact uni? ed parameterized model of facial appearance that takes into account different sources of variability such as individual appearance, 3-D pose, facial expression, change in lighting condition, etc.

The uni? ed approach is very suitable for interpretation and coding of face images. There exist two main phases. r Modeling: In this phase, ? exible models of facial appearance are generated. The shapes of facial features and their spatial relationships are modeled from a set of training images. The model is generated by statistical analysis of the positions of landmark points over the training set in order to describe a mean shape from therein.

This training set is representative of variations due to differences between individuals as well as changes in their expressions and 3-D poses. The shape model is augmented with the gray-level information of the face images. As a result, the shape and gray-level models represent the overall appearance of each face image. This is collectively called the appearance parameter of the face images.

r Interpretation: During this phase, the models are used for coding and interpretation of the face images. When a new query image is presented, the facial features are located automatically based on the ? exible appearance model obtained during the training and modeling. These features are transformed into shape model parameters.The new query face is then deformed to the mean face shape, and the gray-level appearance is transformed into the appearance parameters. Hence, the new face image is coded in terms of these appearance parameters only. This resulting appearance parameters can then be used for personal identi? cation including gender recognition, expression recognition, 3-D pose recovery as well as reconstruction of the face images. Less than 100 parameters are required to describe each image suf? ciently well to generate a good quality reconstruction of the face, inspite of different types of variability.

Experimental results are presented for a database of 690 face images including partially occluded test faces. This uni? d approach is very generic in nature and can be easily adopted for different applications. A similar strategy can also be employed in automatic interpretation of hand gestures, sign reading, lip-reading, etc. [57].

Active appearance model (AAM) is a generative model of a certain visual phenomenon. Although linear in both shape and appearance, yet overall, the AAMs are nonlinear parametric models in terms of the pixel intensities. Fitting an AAM to an image consists of minimizing the error between the input image and the closest model instance; i. e. , solving a nonlinear optimization problem.

Baker et al. have developed facial recognition, interpretation, and coding techniques based on the AAM in [58]–[61]. F.

Gabor Filtering From the standpoint of neurobiology, subtle changes in the shape and texture of the face (that are essential to facial expression discrimination) are best represented using the spatially localized receptive ? elds typical of primary visual cortex cells. Gabor wavelet functions [62] are found to approximately model such behavior [44]. A complex-valued 2-D Gabor function is a plane wave restricted by a Gaussian envelope and is expressed as ? (k, x) = k 2 x2 k2 exp ? 2 ? 2 2? exp(ik, x) ? exp ? ?2 2 . Here the multiplicative factor k 2 ensures that ? lters tuned to different spatial frequency bands have approximately equal energies. The term exp(?? 2 /2) is subtracted to render the ? ters insensitive to the overall level of illumination.

The Gabor wavelet representation allows description of spatial frequency structure in the image, while preserving information about spatial relations. The complex amplitude of the transforms is used as features to test for the presence of spatial structure, restricted to a band of orientations and spatial frequencies, within the Gaussian envelope. The amplitude information degrades gracefully, with shifts in the image location at which it is sampled over the spatial scale of the envelope. Gabor-wavelet-labeled elastic graph matching has been combined with the eigenface algorithm, for facial image 320IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL.

37, NO. 3, MAY 2007 classi? cation [44]. Facial images are transformed using a multiscale, multiorientation set of Gabor ? lters. Rectangular as well as ? ducial (with nodes located at easily identi? able landmarks of the face) grids are registered with the face. This is followed by extraction of complex-valued Gabor transform coef? cients sampled at the grid points and combined into a labeled graph vector. PCA is applied to reduce the dimensionality of the feature space.

Finally, linear discriminant analysis is employed for clustering these Gabor-coded images, based on the different facial attributes.The facial registration grid parameters include the 2-D coordinates of the center of mass, and the horizontal and vertical grid line spacing. The system [44] was implemented on a set of 193 facial images of Japanese females from the Japanese Female Facial Expression (JAFFE) database. 2 Fundamental facial expressions such as being happy, sad, angry, fearful, surprised, and disgusted were classi? ed, along with a “neutral” face.

Each test image was convolved with a set of Gabor ? lters, having highly correlated responses that are redundant at neighboring pixels. A sparse ? ducial grid was positioned by manually clicking on 34 easily identi? able points of each facial image.The regions of the eyes and mouth were found to be the most critical areas for determining facial expressions. The projections of the ? lter responses along discriminant vectors, obtained from the training set, were compared at corresponding spatial frequency, orientation, and locations with the test image, using the normalized dot product to measure the similarity of the two Gabor response vectors. It was observed that the horizontally oriented ? lters were more useful as compared to the vertically oriented ones. This also corroborates with the most noticeable expressive motions of the face such as: 1) opening and closing of mouth and 2) raising and lowering of eyebrows.

The Gabor representation showed a signi? cant degree of psychological plausibility, matching with the semantic rating of images by human observers, and demonstrating a promise for suitable human–computer interface. G. Optical Flow Pixel-based approaches to motion estimation are often referred to as optical ? ow methods [63]. Here, a direct relationship is assumed between object motion and intensity changes within an image sequence. Motion estimation is formulated as an optimization problem where the motion ? eld corresponds to the operator, which best accounts for the intensity variations, given certain restrictions. These methods typically generate dense motion estimates.They perform well when the motion of individual objects is relatively slow, and the scene consists of only a few moving objects. Hence, the dynamics of facial expressions is a good candidate for estimation using optical ? ow.

Optical ? ow has been used to estimate facial muscle action [50], which are triggered by nerve impulses generated by emotions. The muscle actions cause the movement and deformation of facial skin and features such as eyes, mouth, and nose. Since the facial skin at cheek and forehead has a ? ne-grained texture, it is suitable for extracting the optical ? ow. These are 2 http://www.

mis. atr. co. jp/? mlyons/jaffe. html then correlated with different facial expressions. The advantage of the optical ? w algorithm is that there is no need to extract and trace particular points in the image sequence. The subtle texture of the facial skin is suf? cient to extract skin motion in terms of optical ? ow.

In a top–down approach, the optical ? ow ? elds of skin movement is evaluated in muscle windows, each of which de? nes one primary direction of muscle contraction. However, it is necessary to collaborate with psychologists to construct knowledgebases to categorize facial expressions from the muscle movement descriptions. In a bottom–up approach, a 15-dimensional feature vector is used to represent the most active points in terms of the ? rst and second moments of velocity pattern through time at local regions derived from optical ? ow data.Evenly divided rectangular regions are used to cover the entire face as muscle windows, and the most active regions are selected from the sample data to generate the feature vectors.

The k-nearest-neighbors rule [37] is then employed for facial expression recognition. Let us consider an M ? N image sequence with a period of T frames for the kth expression sample. The image is evenly divided into m ? n regions, where r = M/m = N/n. The horizontal and vertical optical ? ow at pixel (x, y) in time frame t are given as ut (x, y) and vt (x, y), respectively. The means of optical ? ow of horizontal and the vertical components at the (i, j)th divided region R(i, j) are expressed as µu,i,j = and µv,i,j = 1 1 T r2 vt (x, y) 0