WIXLIB

International Journal of Engineering Research & Technology (IJERT)ISSN: 2278-0181Vol. 3 Issue 11, November-2014B. Object ClassificationObject classification is the process by which the objectsthat are detected in the frame are classified as what object ofinterest it is. It is essentially just identifying what object it is.Object identification may be done based on differentparameters like shape, motion, colour and texture. So, basedon the parameter used, we can perform shape-basedclassification, motion-based classification, colour-basedclassification or texture based classification.C. Object TrackingObject tracking is a method of following an object throughsuccessive image frames to determine how it is movingrelative to other objects. This is most commonly done bymeasuring the position of the centroid of the object in (x, y) insuccessive frames [19]. Object tracking may be classified aspoint-based tracking, kernel-based tracking or silhouette basedtracking.IV.METHODS OF OBJECT DETECTIONC. Background SubtractionA video sequence may be separated into background andforeground. Foreground usually consists of the objects ofinterest whereas the background data is not important fortracking. If we remove the background data from the videoframe, then we are left with just the necessary data in theforeground, which contains the object of interest.We could get a better accuracy if we already know whatthe background is. For example, in stationary surveillancecameras as in road traffic monitoring, the background isalways constant. The road remains in the same position withrespect to the camera. This gives us the advantage of havingthe background already “modelled” for us. If this is not thecase, then background modelling has to be performed beforebackground subtraction. The objective background modellingis to generate a reference model. The video sequence iscompared with the reference model and the object is detectedby computing the variation between the two. Fig.3. showshow a background is modelled.There are two types ofalgorithms for background subtraction. They are recursivealgorithm and non-recursive algorithm.IJERTObject detection is the first step in tracking an object in thevideo. What actually happens in detection is that “objects” areactually groups of pixels clustered together. The basic idea isto identify these pixel clusters as objects which are not onlymoving in x and y directions but also in time. Let us examineeach of the methods of object detection in detail in thefollowing sections.sequence [20]. Optic flow is a vast area of study and [21]provides a summary of the different methods available forestimation of optical flow. Though this method can get betteraccuracy, it is computationally very costly and its ability todeal with noise is limited.A. Frame DifferencingFrame differencing is basically used to detect movingobjects on a static background. Because objects are movingwith difference to time, the position of the object on one frameis different from the position of the object on the consecutiveframe. By finding the difference between the two frames, wecan get the exact position of the object on the frame. Thecomputational complexity is very low, but complete outline ofthe moving object is difficult to obtain, which also leads tolower accuracy.1) Non - recursive algorithm: A non-recursive techniqueuses a sliding-window approach for background estimation.Select number of previous video frames are stored in a bufferand the background image is estimated based on the temporalvariation of each of the pixels within the buffer. Since onlyselect numbers of frames are stored in the buffer, errorscaused by frames outside the buffer limit are not taken intoconsideration. Storage requirements for non-recursivetechniques may be very large because of large bufferrequirements. This problem is overcome by storing the videoat lower frame rates. Some commonly used non-recursivetechniques include frame differencing, median filtering, linearpredictive filtering and non-parametric modelling [10].Fig. 3. Background modelling (a): Input video frame, (b): Backgroundmodel, (c): Object detected after background subtraction.Fig. 2. Methods of Object DetectionB. Optical FlowOptical flow or optic flow is the pattern of apparentmotion of objects, surfaces, and edges in a visual scenecaused by the relative motion between an observer (an eye ora camera) and the scene [18].Optical flow calculates avelocity for points within the images, and provides anestimation of where points could be in the next imageIJERTV3IS1104582) Recursive algorithm: No buffer is used in the case of arecursive technique. A single background model is updatedbased on each input frame. This means that even frames fromthe distant past could cause an error in the current model.This also reduces the storage space, as no memory would benecessary to buffer the data. An error caused can linger for along time. Some recursive techniques include approximatedmedian filtering, Kalman filtering and Mixture of Gaussians(MoG).www.ijert.org(This work is licensed under a Creative Commons Attribution 4.0 International License.)624

International Journal of Engineering Research & Technology (IJERT)ISSN: 2278-0181Vol. 3 Issue 11, November-2014V.METHODS OF OBJECT CLASSIFICATIONAfter the objects have been detected in a video sequence,the next step would be to identify these objects and classifythem according to our requirement. Classification is donebased on the parameter we select. Depending on whatparameter we select for classification, the methods aredefined as follows: shape-based classification, motion-basedclassification, colour-based classification and texture-basedclassification. Each of the methods is explained below.In all the above mentioned spaces, features to define anobject are not efficient. Therefore in the recent years, colourdescriptors have been classified into histogram based colourdescriptors and SIFT (Scale Invariant Feature Transform)based colour descriptors [12].D. Texture-based classificationTexture is an innate property of virtually all surfaces, thegrain of wood, the weave of fabric, the pattern of crop infields, etc. It contains important information about thestructural arrangement of surfaces and their relationship to thesurrounding environment. Since the textural properties ofimages appear to carry useful information, for discriminatingpurpose features have always been calculated for textures[17].Although it is quite easy for a human observer torecognize and describe in empirical terms, texture has beenextremely adverse to precise definition and analysis bycomputer. Texture is represented by means of texturedescriptors. They observe region homogeneity and histogramsof region borders. Different texture descriptors includehomogeneous texture descriptor (HTD), texture browsingdescriptor (TBD) and edge histogram descriptor (EHD) [18].IJERTA. Shape based classificationShape based classification is done based on shape analysis.Shape analysis is the automatic analysis of geometric shapesby a computer to detect similarly shaped objects by comparingagainst entries on a database. Mostly boundary basedrepresentation is used. However, volume based representationor point based representation of shapes is also possible. Thesimplified representation is called shape descriptor. Acomplete shape descriptor consists of all the informationrequired to reconstruct the shape. Shape descriptors may beinvariant with respect to congruency, isometry (intrinsic shapedescriptors).Graph based descriptors are another class [18].Shapes may also be classified based on part structure. Butcapturing part structure is not a trivial task considering thenon-linearity of shapes. Part structure capturing can basicallybe classified into three categories. The first one builds partmodels from sample images. This requires some priorknowledge of the part. The next two categories capture partstructures from only one image. In the second category,individual parts are compared with each other and a similaritymeasure is obtained. In the third category, the part structure iscaptured considering the interior of shape boundaries [16].reflectance property. Colour information is usuallyrepresented in the most commonly used RGB colour space.The problem with RGB is that it is not a uniform colourspaceTherefore, we have to consider the use of other colourspaces like L*a*b and L*u*v which are perceptually uniform.HSV (Hue, Saturation, and Value) is a relatively uniformcolour space.B. Motion based classificationMotion based classification works on periodicity of themotion. A system can be made to learn how the object movesand then classify it better. Motion based classification has tobe addressed for both rigid and non-rigid objects. Though it iseasier to track objects which are rigid and show periodicity inmotion, a limited amount of periodicity has been known toexist in non-rigid objects as well. Optical flow is also used formotion-based classification [10].Fig.4. Methods of Object ClassificationC. Colour-based classificationColour-based approach is based on studying the colourfeatures in an image. The two main colour features that areused to classify based on colour are spectral powerdistribution of the illuminant and the object‟s surfaceIJERTV3IS110458VI.METHODS OF OBJECT TRACKINGHaving detected the objects and classified them, the nextstep would be the actual tracking process. According to [21],tracking can be defined as the problem of estimating thetrajectory of an object in the image plane as it moves around ascene. There are three methods of object tracking which arediscussed in detail below. They include point tracking, kerneltracking and silhouette tracking [18].A. Point TrackingIn the point tracking approach objects are represented aspoints and are generally tracked across frames by evolvingtheir state (object position and motion). Point tracking may beKalman filtering, particle filtering or Multiple HypothesisTracking (MHT).1) Kalman Filtering: Kalman filtering is an algorithmthat uses a series of measurements observed over time,containing noise (random variations) and other inaccuracies,and produces estimates of unknown variables that tend to bemore precise than those based on a single measurement alone.More formally, the Kalman filter operates recursively onstreams of noisy input data to produce a statistically optimalestimate of the underlying system state [14].There are two steps in the algorithm. The prediction stepproduces estimates of the current state variables along withtheir uncertainties. Then, the outcome of the nextmeasurementwww.ijert.org(This work is licensed under a Creative Commons Attribution 4.0 International License.)625

International Journal of Engineering Research & Technology (IJERT)ISSN: 2278-0181Vol. 3 Issue 11, November-2014Fig. 5. Methods of Object Tracking4) matrix must also be defined. Each element of thematrix, cij, represents the probability that measurement i hasbeen originated due to j. MTH is capable of tracking multipletargets, handling occlusions and calculating optimal solutions[15].B. ernel Based TrackingIn kernel based approach, an object is tracked based oncomputing the motion of the rectangular or elliptical kernel inconsecutive frames. Motion may include translation, rotationand affine transformations.IJERTis observed and the estimates are updated using weightedaverage, with more weight being given to estimates withhigher certainty. Since it is a recursive algorithm, only thepresent value, previous value and uncertainty matrix areenough to calculate in real time.2) Particle Filtering: Particle filters are a set of on-lineposterior density estimation algorithms that estimate theposterior density of the state-space by directly implementingthe Bayesian recursion equations. Posterior density isrepresented by a set of particles. No assumptions about thedynamics of the state-space or the density function are made,but they provide a well-established method for generatingsamples from the required distribution. The samples arerepresented by a set of particles and each particle has aweight assigned to it. The weight of each particle representsthe probability that that particle is being sampled from theprobability density function. Resampling is done so as toavoid weight disparity which leads to weight collapse. Whenthe state varibles are not distributed normally (Gaussian),Kalman filter provides a poor approximation. Particle filtersare used to overcome this problem. This algorithm usescontours, colour features or texture mapping [18].3) Multiple Hypothesis Tracking(MHT): The MHTalgorithm begins with the set of hypothesis of the previousiteration also called parent hypothesis set and the set ofmeasurements from the beginning until that iteration. Eachhypothesis represents a different set of assignments of the setof measurements to the different tracks. Taking into accountthe new set of measurements and one of the previoushypotheses, a new hypothesis is generated, making a specificassignment of the current measurements. The set of plausibleassignments that can be done for a parent hypothesis isnamed ambiguity matrix (sometimes also called hypothesismatrix). Each element of the matrix, aij, can take a value of 1or 0, representing the possibility that measurement i isassociated to a previous track, a new track, is considerednoise, etc., or not. Associated to the ambiguity matrix, a costIJERTV3IS110458The problem with this method is that a part of the objectmay be left outside the kernel while a part of the background,which is not necessary, may be added into the kernel.Tracking may be based on geometric shape of the object,object features and appearance. Different kernel basedtracking approaches include simple template matching, meanshift method, support vector machine (SVM) and layeringbased technique. These are explained below.1) Simple Template Matching: Template matching is atechnique used in digital image processing for finding smallparts of an image or video that match a template image. Sincethere is no specific way of right and wrong, it is termed to bea brute force method of examining regions of interest in avideo. The matching procedure calculates a numerical indexfor how well the image in the frame matches with the imagein the template.Motion based classification works on periodicity of themotion. A system can be made to learn how the object movesand then classify it better. Motion based classification has tobe addressed for both rigid and non-rigid objects. Though it iseasier to track objects which are rigid and show periodicity inmotion, a limited amount of periodicity has been known toexist in non-rigid objects as well. Optical flow is also used formotion-based classification.2) Mean Shift Method: Mean shift algorithm is a nonparametric feature space analysis technique for locating themaxima of a density function [18]. It is an iterative algorithmand is based on predicting the future values based on the pastvalues. One of the simplest forms would be to calculate theconfidence map of the new image based on colour histogramwww.ijert.org(This work is licensed under a Creative Commons Attribution 4.0 International License.)626

International Journal of Engineering Research & Technology (IJERT)ISSN: 2278-0181Vol. 3 Issue 11, November-2014the contour of the next frame. Only if this requirement issatisfied, proper tracking results can be achieved.There are two ways to perform contour tracking. In thefirst approach, the contour shape and motion are modeledusing state space models. The second method is more directand direct minimization techniques such as gradient descentfor minimizing the contour energy, thereby evolving thecontour. The most significant advantage is that this can beused to track objects of irregular shapes.2) Shape matching: Shape matching is similar to theshape-based classification technique. But, instead ofclassifying the object, we use shapes to track the particularshape. It is also similar to template matching, because theshapes are stored in a database and then the shape on theframe is compared with the shape in the database and thustracking is done. Silhouettes from two successive frames canalso be matched to obtain the required results. A single objectcan be tracked and occlusions are handled by means of usingHough transform [11].VII. CONCLUSIONWe have surveyed the various steps involved in trackingan object from a video sequence and also the differentmethods to perform object detection, object classification andobject tracking were discussed. The advantages anddisadvantages of each of these methods were discussed. Inthe future, we plan to propose an algorithm that overcomesthe disadvantages of the existing object detection methodsand track objects in a video with capability to handle multipleobjects and occlusions.IJERTof the previous images. The peak of the confidence map willoccur at the next predicted position of the object. Thus, theconfidence map is a probability density function and the peakof the function may be calculated using mean shift method[15].The region of interest (ROI) is selected by a bounding boxfrom the first frame of the video. The probability densityfunction is calculated based on colour information. Then, thepresent probability density function is compared with theprobability density functions of the consecutive frames.Degree of similarity between the frames is represented byusing Battacharya coeffecient. The same procedure will befollowed till the last frame of the video is reached [11].This method has a lot of drawbacks. Only one object canbe tracked. The ROI has to be initialized manually. It cannottrack an object if the object is moving with high speed withinthe frame.3) Support Vector Machine (SVM): Support VectorMachine uses algorithms which are supervised learningmodels and their associated learning algorithms whichanalyze data and recognize patterns based on previouslylearnt data. Support vector machines are used mostly forclassification and regression [18].The basic idea behind a SVM algorithm is that it classifiesthe points into two sets of hyperplanes

The pros and cons of each of the methods are discussed. Object detection methods are frame differencing, optical flow and background subtraction. Then, objects may be classified based on shape, motion, colour and texture. Tracking methods involve point based tracking, kernel based tracking and silhouette based tracking.