1.
Shivganga Udhan, Prof. Pravin Futane / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 1, January-February 2013, pp.289-293 Alphabet Recognition Using Hand Motion Track Shivganga Udhan* , Prof. Pravin Futane**. ME student Sinhgad College of Engineering Pune*, Professor and HOD Sinhgad College of Engineering Pune**.Abstract This paper describes method for interface to browse and edit photos.[5]Cha-Suprecognition of alphabets from hand motion Jeong,Dong-Seok Jeong,Designed method based ontrajectory. This method uses Hidden Markov contour information and fourier descriptors [9].Model (HMM). Gesture recognition for Hongwei Ying, proposed a method for the trackingalphabets is done in three main stages; of fingertip in scenes with complex backgroundpreprocessing, feature extraction and [10]. Jonathan Alon, Quan Yuan,introduces aclassification. In first stage, preprocessing hand Unified Framework For simultaneously performingis detected using color information. After spatial segmentation, temporal segmentation anddetection of hand, motion trajectory which is also recognition[11].called as gesture path will be determined bytracking the hand. The second stage, feature II. LITERATURE SURVEYextraction gives pure path by enhancing gesture In earlier work, the effects of varyingpath it also determines the orientation between weighting factors in learning from multiplethe center of gravity and each point in gesture observation sequences is studied. Multiple-sequencepath. This orientation vector gives discrete vector Baum-Welch, Ensemble methods, and Viterbi Paththat is used as input to HMM. In the final stage Counting methods are compared on Synthetic andalphabet is recognized by gesture path. Our real data. This study determined that Viterbi Pathmethod will recognizes alphabets and we are Counting was the best and fastest method. Howeverexpecting more than 90% of recognition rate. the problem of choosing the initial model for the training (re-estimation) process was discussed by MKeywordsβHand Tracking, Hidden Markov Elmezain, A. Al-Hamagi,G. Krell, S. El-Etriby, B.Model, Human Computer Interaction, Gesture Michaelis[1].Recognition A method to determine an alphabet from a single hand motion using HMM is described by MI. INTRODUCTION Elmezain, A. Al-Hamagi,G. Krell, S. El-Etriby, B. Human-machine interfaces are becoming Michaelis[1].System overview of their work is givenmore important as computer technology is growing. as follows:Now a days, hand gesture have been receiving a lotof attention in the area of HCI as an alternative toconventional interface devices. The hand provides anatural means for communication, which is highlyexpressive and simple to manipulate. Now Currentfiled of research is Hand gesture recognition and alot of work is being done on it. There are so manyapplications have been designed on it. Computerversion hand based tracking is used to interact withcomputers. method can take benefit by using theHidden Markov Model(HMM).This method isintroduced many days before, but it is very popularnow, due to the effective use in the area of gesturerecognition.Some kind of Artificial intelligence isrequired for recognising the things from the gesturepath or motion trajectory. Barbara Resch have given a gentle Fig 1. Recognition of Alphabet using HMMintroduction to Markov models and hiddenMarkov (Courtesy from [1]).models (HMMs) and relates them to their use inautomatic speech recognition. [7].Nianjun Liu have 1. Preprocessing: point out hand location and trackgiven several ways to initialize and train hidden hand motion to generate its motion trajectorymarkov model (HMMs) for gesture recognition[8]. (gesture path).Ankit Gupta, Kumar Ashis Pati used hand trakingand finger tip detection to create a simple user 289 | P a g e
2.
Shivganga Udhan, Prof. Pravin Futane / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 1, January-February 2013, pp.289-2932. Feature Extraction: enhance the gesture path to determined by Eq. 1 and Eq. 2. Since the location of give pure path and then quantize the orientation pure path for the same gesture according to the start to determine the discrete vector point is different, we calculate the orientation3. Classification: the graphical gesture path is between any point in pure path and center of gravity recognized using discrete vector and Left-Right (fig.3). banded model with six states.A. Preprocessing First stage in our method is preprocessingwhich will be divided into two steps. The first stepis skin color detection and second step is handlocalization and tracking.B. Feature extraction Fig. 3: Orientation and vector quantization range The feature extraction is a very important (Courtesy from [1]).part in this method to recognize the alphabet gesturepath. There are three basic features as location,orientation and velocity. The previous researches π 1showed that the using of orientation feature is the ππ = ππ‘ (1)best in terms of results. So, it is more reliable. A π π‘=1gesture path is spatio-temporal pattern which π 1consists of centroid points (Xhand, Yhand). The ππ = ππ‘ (2)gesture path is having some stable (i.e. non π π‘=1movable) points possibly starting and end point of ππ β π¦ π‘ π π‘ = ππππ‘ππ ; π‘ = 1,2, β¦ , π (3)gesture path, so enhance the gesture path to obtain a ππ β π₯π‘pure path as follows (Fig2. ): Where n is the length of pure path. In addition, the orientation ΞΈt determined according to eq. 3 where the orientation is divided by 30 in order to quantize the value of it from 1 to 12. Thereby, the discrete vector is obtained which is used as input to HMM. C. Classification Classification is the last stage in this method. Throughout this step, the pure path of hand graphical is recognized by using Left-Right Banded model with 6 states and building gesture database. According to this stage, Baum-Welch algorithm (BW) is used for training the initialized parametersFig2. Gesture Path Enhancement (Courtesy from of HMM to provide the trained parameters. Then the[1]). trained parameters and discrete vector are used as input to Viterbi algorithm in order to obtain the best Firstly, the gesture path is entered to a path. By this best path and gesture database, thecheck start state to see, whether there is no motion if pure path is recognized. The following subsectionsyes, then it takes the next point else the pure path is describe this stage in detail [1].begin generated. Then the pure path state 1) Hidden Markov Models: Markov model is acontinuously generates the pure path while the mathematical model process where thesepoints of gesture path input move. When the point processes generate a random sequence ofdoes not move, check end state is called. Finally, the outcomes according to certain probabilities.pure path is generated from check end state when An HMM is a triple A = (A, B, Ξ ) asinput gesture path is ended. Also while there are follows:points in gesture path, this state perform a check to ο· The set of states S {s1, s2, , SN} where Nsee if the point not moves then delete it and takes is a number of states.the next point, else return to the pure path state. ο· An initial probability for each state Ξ i, i=l, In contrast to the enhancement gesture path 2,....N such that Ξ i = P(si) at the initial step.to obtain a pure path, our method is based on the ο· An N-by-N transition matrix A = {a ij}angle(orientation) as a basic feature. Therefore, the where aij is the probability of a transition fromorientation is based on (Xc,Yc) where Xc and Yc state Si to Sj; 1 <=i, j<=N and the sum of theare the center of gravity of pure path and are entries in each row of matrix A must be 1 290 | P a g e
3.
Shivganga Udhan, Prof. Pravin Futane / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 1, January-February 2013, pp.289-293 because this is the sum of the probabilities of duration time d of states for each alphabet such that making a transition from a given state to each d is defined as; of the other states. π π= ο· The set of possible emission (an π observation) 0={o1,o2, . . on} where T is the Where T is the length of pure path and N length of pure path. represents the number of states that has a value 6 in ο· The set of discrete symbols V {vl, v2, ..., this method. vI} M represents the number of discrete π΄ symbols. π ππ 1 β π ππ 0 0 0 0 ο· An N-by-M observation matrix B ={bim} 0 π ππ 1 β π ππ 0 0 0 where bim gives the probability of emitting 0 0 π ππ 1 β π ππ 0 0 = symbol Vm from state s1 and the sum of entries 0 0 0 π ππ 1 β π ππ 0 in each row of matrix B must be 1 because this 0 0 0 0 π ππ 1 β π ππ is the sum of the probabilities of making a 0 π ππ 1 β π ππ 0 0 1 transition from a given state to each of the other states. Such that; There are three main problems for HMM:Evaluation, Decoding and Training that can be 1 π ππ = 1 βsolved by using Forward- Backward algorithm, πViterbi algorithm and Baum-Welch algorithmrespectively[1]. Also, HMM has a three topology: The second important parameter is a matrixFully Connected (Ergodic model) where any state in B that is determined by Eq. 7. Since HMM states areit can be reached from any other state, Left-Right discrete, all elements of matrix B can be initializedmodel such that each state can go back to itself or to with the same value for all different states.the following states and Left-Right Banded (LRB)model that also each state can go back to itself or the 1 π΅ = π ππ }; π ππ = (7)following state only (Fig 4.). π Where i, m run over the number of states and the number of discrete symbols respectively. The third parameter in the HMM is the initial vector Ξ which takes value;Fig 4. : Left-Right Banded model with 6 states π = 1 0 0 0 0 0) π(Courtesy from [1]). That is because we use 6 states as the2) Initializing parameters for LRB model: It maximum numbers of the segmented graphicalwill be more convenient to explain, why we use left- alphabet and in order to guarantee that it begin fromRight Banded model with 6 states before describing the first state as shown in Fig. 6.the initialization of HMM parameters. Since eachstate in fully connected model has many transitions 3) Baum-Welch and Viterbi Algorithm.rather than LRB model, the structure data can be After the HMM parameters are initialized,losing easily. Moreover, LRB model is restricted M Elmezain, A. Al-Hamagi,G. Krell, S. El-Etriby,and simple for training data that will be able to B. Michaelis[1], used Baum-Welch algorithm tomatch the data to the model. In addition, the discrete perform the training for initialized parameters wherevector contains a single sequence of codehook from the inputs of this algorithm are discrete vector i.e.1 to 2 in this method. For that reason the LRB obtained from feature extraction stage and initializemodel is preferred than left-right model. About the parameter. This algorithm give us a new parameternumber of steps considers the number of segmented estimation of vector Ξ , matrix A and matrix B. Inpart that is contained in graphical pattern when we the next step the viterbi algorithm takes the discreterepresent it. For example, βLβgraphical pattern vector, new matrix A and new matrix B as its inputcontains two segmented parts, thus we need only 2 and gives us the best path. For doing that the initialstate for it, while βGβ and βEβ patterns need 5 and 6 state determined by product initial vector Ξ withstates respectively. Therefore the no of states is 6 associated observation probability bit. After this thenearly for all alphabets. For this reason we selected best result route of the next step (t+1) is determinedthe left-right banded model with 6 states. There is is by taking the minimum probability that is derivedno doubt that, a good parameters initialization for from the product of previous state observationHMM (A, B, Ξ ) gives a better results. The matrix A probability with its transition. Finally, byis determined by Eq. 5 and it depends on the backtracking through the trellis, the best path is obtained by selecting the maximum probability state 291 | P a g e
4.
Shivganga Udhan, Prof. Pravin Futane / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 1, January-February 2013, pp.289-293at time T as shown in fig 7. After the best path is find the centroid if actual skin part (i.e. users hand)determined, we call the database gesture for this for each frame of input video. Then these centroidspath to recognize it from A to Z by using higher are joined to get gesture path (i.e. motion trajectory).priority for comparing and choosing. The following After that we use Left-Right Banded model onsteps demonstrate how Viterbi algorithm works: gesture path to get pure path and thus build the gesture database.1) Initialization:For 1 β€ π β€ π, IV. RESULTS AND DISCUSSIONS a) πΏ1 π = π π . π π (π1 ), These are some results which we got till now are b) β π (π) = 0, shown in fig.6 below. Here are first and last frames i.e. starting and end points of gesture path for2) Recursion: For 2 β€ π‘ β€ π, 1 β€ π β€ π, alphabet βAβ, and centroids of hand in respective a) πΏ π‘ π = max π πΏ π‘β1 π . π’ ππ . ππ π£ π‘ frames. b) β π‘ (π) = πππ max π πΏ π‘β1 π . π ππ3) Termination: a) πβ = max π [πΏ π (π)] b) π βπ = πππ max π [πΏ π (π)]4) Reconstruction: For t=T-1, T-2, β¦, 1. π βπ = β π‘+1 (π β ) π‘+1III. SUGGESTED METHODAlphabets recognition process mainly consists ofthree stages as given in fig. 5 below: Fig. 6: Start and end point for βAβ and their centroids. Possible gesture paths (Motion trajectory) for all 26 alphabets are as shown in fig.7 below. That means we are not considering the letter as it is but we can skip some part without any problem to reduce complexity of overlapping of gesture path. For example we can skip middle horizontal line from βAβ as shown in figure.Fig. 5: System Overview. The proposed system is self contained product Fig. 7: Letter Gesture forthis is basically a means of communication between A,B,C,D,E,F,T,U,V,X,Y,Z.physically disable people with computer or it can bereplacement for input devices of computer systems. V. SUMMERY AND CONCLUSIONMain goal of this system is to accept the video of This paper presents a method for vision-hand motion and to display appropriate alphabet based recognition of alphabets from A to Z by usingidentified from hand motion trajectory. In this Hand Motion Trajectory. It uses HMM formethod the gesture path is generated from hand recognizing Alphabet. This method mainly consistsmotion trajectory by using HMM. This gesture path of 3 main stages. The first stage is the preprocessingwill recognise respective alphabets. where hand is localized and tracked to produce In our method we are hiding head of user gesture path. In the second stage i.e. featureform camera to reduce the complexity of the Extraction, where the quantization and orientation isSegmentation and Localization module. Then we used to get the discrete vector. The final stage is 292 | P a g e
5.
Shivganga Udhan, Prof. Pravin Futane / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 3, Issue 1, January-February 2013, pp.289-293classification, in this stage actual recognition of Segmentationβ,2009 IEEE Computer Soalphabet is done. ciety Pp1685-1699. [12] Ming-Hsuan Yang and Narendra AhujaACKNOWLEDGEMENT βRecognizing Hand Gesture Using Motion The research in this paper was carried out Trajectoriesβ Computer Vision and Patternat Sinhgad College of Engineering, Pune. So, Recognition, IEEE Computer Societyspecial thanks to Head of Computer Department, Conference on Jun 2002.Principal and Management of Sinhgad College of [13] Nianjun Liu Brian et.al βEvaluation ofEngineering. HMM Training Algorithms for Letter Hand Gesture Recognitionβ Intalligent Real-REFERENCES Time Imaging and Sensing (IRIS) [1] M Elmezain, A. Al-Hamagi,G. Krell, S. El- Group.Pp 648-651. Etriby, B. Michaelis, βGesture Recognition for alphabet from Hand Motion Trajectory AUTHORS Using Hidden Markov Modelsβ. 2007 First Author β Ms. Shivganga Udhan, ME Computer IEEE International pp:1192-1197. Networks Student, Sinhgad College of Engineering, [2] N. Liu, B. C. Lovell, P. J. Kootsookos, R. I. Pune, A. Davis, W. 2004 βModel Structure Selection and Training Algorithm for a Second Author β Prof. P. R. Futane, Head of HMM Gesture Recognition Systemβ, Computer Department, Sinhgad College of International Workshop in Frontiers of Engineering, Pune, Handwriting Recognition, pp.-100-106, October 2004. [3] Lawrence R. Rabiner, Fellow, βA Tutorial on Hidden Markov Model and Selected Applications in Speech Recognitionβ Proceedings of the IEEE, vol. 77, No. 2, Feb 1989. Pp 257-286. [4] Cha-Sup Jeong, Dong-Seok Jeong βHand- written Digit Recognition Using Fourier Discriptors and Contour Informationβ 1999 IEEE TENCON. [5] Ankit Gupta, Kumar Ashis Pati βFinger Tips Detection and Gesture Recognitionβ Indian Institute of Technology, Kanpur,India. November 12, 2009. [6] Ming-Hsuan Yang and Narendra Ahuja βRecognizing Hand Gesture Using Motion Trajectoriesβ Computer Vision and Pattern Recognition, IEEE Computer Society Conference on Jun 2002 [7] Barbara Resch , Erhard and Car line Rank βHidden Markov Modelβ. Signal Processing and Speech Communication Laboratory. [8] Nianjun Liu, Richard I.A Davis,Brian C.Lovell and Peter J.Kootsookos βEffect of initial HMM Choice In multiple sequence training for gesture recognitionβ IEEE. [9] Cha-Sup Jeong, Dong-Seok Jeong, 1999 βHand-Written digit recognition using fourier descriptors and Contour informationβ, IEEE TENCON. [10] Hongwei Ying,Jiatao Song,Xiaobo Ren and Wei Wang βfingertip detection and traking using 2D and 3D informationβ, 2008 IEEE Pp1149-1152 [11] Jonathan Alon, Quan Yuan βA Unified Framework for Gesture Recognition and Spatiotemporal Gesture 293 | P a g e
Clipping is a handy way to collect and organize the most important slides from a presentation. You can keep your great finds in clipboards organized around topics.
Be the first to comment