Published on

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

Published in: Technology, Business
  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. Gopika Mane, Krunank Panchal, Sanchita Sable / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.292-297292 | P a g eAnalysis Of Various Techniques Used For Implementation OfVideo Surveillance SystemGopika Mane*, Krunank Panchal**, Sanchita Sable****(Marathwada Mitra Mandal‟s Institute Of Technology (MMIT), Lohgaon, PUNE University, Pune-47)** (Department of Information Technology (MMIT), PUNE University, Pune-47)*** (Department of Information Technology (MMIT), PUNE University, Pune-47)ABSTRACTVideo Surveillance software’s aredeveloped in order to provide the user of thesoftware security from threats to data and otherproperty from burglars. Image and Pixelmatching algorithm are required forimplementing the Video Surveillance System.This is especially required for computer securityapplications, where our algorithms areeffectively used. These algorithms are useful fortracking an object in motion and classifying it onthe bases of motion level, which would help insubsequent motion detection analysis. So thatwhen an activity occurs in absence of user, theapplication detects it and performs the requiredaction. This algorithm helps to representgraphically the amount of changes occurred inthe environment.Keywords - Cluster Matching, ImageComparisons, Pixel Comparisons, Pixel levelEstimation, Similarity Threshold, and Sub-pixellevel Estimation.1. INTRODUCTIONVideo surveillance systems produce hugeamounts of data for storage and display. Long-termhuman monitoring of the acquired video isimpractical and ineffective. Automatic abnormalmotion detection system that can effectively attractoperator attention and trigger recording is consideredas the key to successful video surveillance indynamic scenes, such as airport terminals. Thevideo-surveillance architectures are used withlimited computing power and available near thecamera for compression and communication. Thealgorithm uses the macro block motion vectors thatare generated in any case as part of the videocompression process. Motion features are derivedfrom the motion vectors.The algorithm is modular, in the sense thatdifferent feature vectors are suggested andalternative probability density estimation ormodeling methods are be used. The input to thealgorithm is the set of macro-block motion vectors(such as intra-frame and intra-block flags) that areproduced anyway by the compression proces whichis an essential part of many modern videosurveillance systems. The algorithm is used mainlyfor triggering video recording for later humananalysis and then transmission to a human observerby detecting the „object‟ that generated theabnormal motion.In this paper, we represented an algorithmwhich works with the idea of building a model ofthe static scene (i.e. without moving objects) calledbackground, and compares every frame of thesequence to this background with respect todiscriminate the regions of abnormal motion,called foreground (the moving objects). Thereforewe are interested with the algorithm that must workwithout or new static objects settling in the scenewhich means that the background must be kepttemporally adaptive, so that there must be a localestimation for the background value, and thesealgorithms do not use much resource likecomputing power and memory. These usedalgorithms imply that the statistical measures on thetemporal activity must be kept locally available inevery pixel, and which needs to constantly updated.This algorithmic approach makes use of the singlemodel like the previous frame or a temporalaverage for the background, and global thresholdfor decision. Also there are some backgroundestimation methods that are based on the analysis ofthe histogram check of the values that are taken byeach pixel within a fixed number of past frames,with the features like mean, median or mode of thehistogram can be chosen to set the backgroundvalue, and the foreground that can be discriminatedby comparing the difference between the currentframe and the background with the histogram checkvariance. Therefore the video surveillance systemrequires the recursive methods that do not keep inmemory a histogram for each pixel, but rather afixed number of estimates are computedrecursively. This value of the variance is useddirectly afterward for the detection of moving area.2. LITERATURE SURVEY OFALGORITHMS:2.1 Pixel Matching AlgorithmPixel matching method takes the input ashigh dynamic range image that maps it to a limitedrange of luminance values reproducible with thedisplay device (i.e. video surveillance system). Theapproach follows functionality of attempting to
  2. 2. Gopika Mane, Krunank Panchal, Sanchita Sable / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.292-297293 | P a g econstruct a sophisticated model. The operationrequires to be performed in three steps. In the firststep, we need to estimate local adaptationluminance in the image at each point. After that insecond step, the values are applied to the simplefunction so as to compress those values into therequired display range. As there is possibility thatthe details of the important image can be lostduring this process, and solution need to re-introduce details over the image in the final pass. Ifprovided with an accurate description of theenvironment, these methods are capable ofcomputing luminance values at each pixel whichclosely match those measured at correspondingpoints in the real scene.2.1.1 Agglomerative Clustering Algorithm:Agglomerative clustering builds thesolution by initially assigning each point to its owncluster and then repeatedly selecting and mergingpairs of clusters. Thus, it builds a hierarchicalmerging tree from the bottom (leaves) towards thetop (root). The key parameter here is the criterionused for selecting clusters to be merged. We focuson the Group Average criterion, which measuresthe similarity of two candidate clusters as theaverage pair wise similarity between theirmembers. Thus, the average-link criterion allowsspecifying the size or compactness of the resultingclusters. This property is very useful in buildingcompact appearance clusters and makes thealgorithm robust to outliers. A similarity thresholdthat produces visually compact Clusters onlydepend on the employed feature descriptors, thuscan be estimated experimentally and used ondifferent data sets. Another advantage ofagglomerative methods is that given the clusteringtrace from a full hierarchical clustering, i.e. theindices of clusters merged in every step and thesimilarities between them, we can rebuild theclusters for a different similarity threshold at almostno computational cost.The Algorithm is basic straightforward1. Compute the proximity matrix2. Let each data point be a cluster3. Repeat4. Closest two clusters are merged.5. The proximity matrix needs to update6. Until one single cluster remainsThe main drawback of theagglomerative clustering algorithm is its O (N2 logN) run-time and O (N2) space complexity. Thiscomes from the requirement that clusters should bemerged in decreasing order of similarity and thatthe distances must be recomputed after eachagglomeration. In order to make agglomerativeclustering applicable to large data sets, bothcomplexities have to be reduced. [5][7]2.1.2 Reciprocal nearest Neighbor PairAlgorithm:The improved clustering method is based on theconstruction of reciprocal nearest neighbor pairs(RNN pairs), that is of pairs of points a and b, suchthat a is b‟s nearest neighbor and vice versa. RNNis applicable to clustering criteria that fulfillreducibility propertyd(ci,cj) ≤ inf(d(ci, ck),d(cj ,ck))⇒inf(d(ci,ck),d(cj,ck)) ≤ d(ci ∪cj ,ck)An amortized analysis shows that this algorithmhas a computational complexity of O (N2d) withonly linear space requirements. The timecomplexity is high when N is large. Here wepresent a strategy to further improve also the run-time efficiency.[5][7]2.1.3 Sub Pixel Mapping Algorithm:This algorithm creates problem of accuracy when itundergoes mixed pixel in remote sensingclassification. This algorithm consists of twostages:1. Pixel level Corresponding Estimation:In Pixel level Estimation, We detect correspondingpoint q0 with pixel level accuracy.2. Sub Pixel level Corresponding Estimation:In Sub Pixel level Corresponding Estimation, Werecursively improve the sub pixel accuracy byadjusting location of sub pixel windowalignment.[9][10]Procedure for sub pixel level Estimation:Input: Images I0(n1,n2) and J0(n1,n2), Referencepoint p0 in I0(n1,n2) given in pixel level accuracy,corresponding point q0 in J0(n1,n2) given in pixellevel accuracy,Output: Reference point p-1 in I0(n1,n2) given withpixel level accuracy. Corresponding point q0 inJ0(n1,n2) given with pixel level accuracy.Procedure for pixel level Estimation:Input: Images I0(n1,n2) and J0(n1,n2), Referencepoint p0 in I0(n1,n2)Output: corresponding point q0 in J0 (n1,n2)
  3. 3. Gopika Mane, Krunank Panchal, Sanchita Sable / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.292-297294 | P a g e2.1.4 Patch Matching CorrespondingAlgorithm:Patch Match is one of the fast algorithmsthat is used to compute the approximate dense K-nearest neighbor correspondences between blotchesof two image region. Patch Match is evaluated inthree ways: (1) to find k nearest neighbors,(2) tosearch across rotations and scales, just for additiontranslations, and (3) to match distances usingarbitrary descriptors, and not just with sum-of-squared-differences on patch colors. The PatchMatch algorithm finds dense, globalcorrespondences an order of magnitude faster thanprevious approaches, such as dimensionalityreduction (e.g. PCA) combined with tree structureslike kd-trees, VP-trees, and TSVQ.The algorithmfinds an approximate nearest-neighbor in an imagefor every small (e.g. 9x9) rectangular patch inanother image, using a randomized cooperative hillclimbing strategy. Here, the basic algorithm findsonly a single nearest-neighbor, at the same scaleand rotation. The core algorithm is generalized andextended.First, for problems such as object detection,denoising, and symmetry detection, we can detectmultiple candidate matches for each query patch.Thus we extend the core matching algorithm to findk-nearest neighbors (k-NN) instead of only 1-NN.Second, for problems such as super-resolution,object detection, image classification, and tracking(at re-initialization), the inputs may be at differentscales and rotations, therefore, it should extend thematching algorithm to search across thesedimensions. Third, for problems such as objectrecognition, patches are insufficiently robust tochanges in appearance and geometry, so we showthat arbitrary image descriptors can be matchedinstead[4][5]2.1.5 Tone Mapping Algorithm:Tone mapping techniques that is used forimage processing and computer graphics needs tomap one set colors to another color set so as toorder approximate the appearance of high dynamicrange images approximate in the medium that has adynamic limited range. All Print-outs, CRT or LCDmonitors, and projectors have the dynamic limitedrange which is inadequate and reproduce the fullrange of intensities light which is present in naturalscenes. Tone mapping suffers the problem with thereduction strong contrast received from theradiance scene to the range displayable that ispreserving the details of image and appearance ofthe color that is important to appreciate the contentof the original scene.[11]Fig 1 Tone mapping in digital photography2.1.6 K-Means Algorithm:It is Partitioned clustering approach where eachcluster gets associated with a center point(centroid). Cluster is assigned with each point thatis the closest centroid .K needs to be specified asthe number of clusters,The basic algorithm is very simple1: k points need to be selected as the initialcentroids.2: Repeat loop3:Create k clusters by assigning points to theclosest centroid.4: Recomputed the centroid of each cluster.5: Until the centroids don‟t change.Mathematical module for K-MeansK-means is an algorithm used for partitioning (orclustering) N data points into K disjoint subsets Sjthat contains Nj data points which are required tominimize the sum-of-squares criterionWhere xn is a vector representing the nth data pointand mj is the geometric centroid of the data pointsin SjComplexity is O( n * K * I * d )n=number of points,K = number of clusters,I = number of iterations,d = number of attributesFig 2 K-Means Algorithm2.2 Image Matching Algorithm:Image matching is an importantfundamental task in a variety of image processingapplications as motion analysis, image sequenceanalysis, stereo vision. These matching methodscan estimate the displacement between two imagesand employs (1) an analytical function fittingtechnique to estimate the position of the correlationpeak, (2) a windowing technique to eliminate theeffect of periodicity, (3) a spectrum weighting21|| jKj SnnxJ   
  4. 4. Gopika Mane, Krunank Panchal, Sanchita Sable / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.292-297295 | P a g etechnique to reduce the effect of aliasing and noise.2.2.1 Classic image check:The classic image check is used to compare thevalue of color of every single expected pixel andactual pixel. As one of the expected pixels getsdiffered from the actual pixel then the check fails.The Tolerance option is required for checkingimages and then defining the tolerance required forcomparing pixel values. This pixel based check isused and suitable if we want an exact match imagethat has tolerances minimal or any deviations. If theapplication renders as the component is not fullydeterministic, then this algorithm is not appropriate.Example-The classic image check does not alter theimage, and therefore the result looks identical asper the original image.Fig 3 Classic image check algorithmThe classic image check is used for the Algorithmsimage comparison attribute when empty.2.2.2 Pixel based identity check:This algorithm is similar to the classicimage check algorithm, but it requires accepting theamount of pixels unexpected. Every pixel is split into its three sub-pixels red, green and blue. Afterthat it requires to checks every actual color valuecompared to the expected color value. The finalresult contains the amount of pixel identical anddivided with the total pixel amount. The outputcalculated is checked with the expected value.When images are not rendered as fullydeterministic but we need to accept unexpectedpixels with certain percentage, then this algorithmmay be used. It is not used, when the actual imagesare used to have distortions or shifts.Example-The resultant image of the typicalalgorithmAlgorithm=identity; Expected=0.95;looks same as per the unique image as thisalgorithm does not consider image manipulation.Fig 4 Pixel based identity check algorithmParametersAlgorithm=identityShould use the Pixel based identity check .ExpectedExpected probability needs to be defined which isvalid between values 0.0 and 1.0.Resize (optional)Needs to be defining when the actual image needsto be resized before calculation to match theexpected image size. Valid values may be "true"and "false".Find (optional)Required when an image-in-image search.2.2.3 Pixel based similarity check:This algorithm needs to be split in everypixel through its three sub-pixels which are red,green and blue. After that it checks every actualvalue color with the expected value color tocalculate the similarity percentile. All thedeviations percentile are added up and used tocompute the ultimate result with the averagedeviation against all color values and all pixelsvalues. The calculated result is compared with anexpected value. When images are not rendereddeterministic fully but we need to accept a certaindeviation, then this algorithm can be used to fulfillthis requirement. When we accept deviations forsome pixels, the deviation average for an image issmall, and this algorithm can be used to fulfill thepurpose. It is not used, when the actual images areneeded to shifts or distortions.Example-The resultant image of the algorithmexemplary:Algorithm=similarity; Expected=0.95; Looksidentical as per the original image as this algorithmdoes not follow image manipulation.Fig 5 Pixel based similarity check algorithmParametersAlgorithm=similarityShould use the Pixel based similarity checkExpectedExpected probability needs to be defined and validvalues are between (0.0 and 1.0).Resize (optional)Required when the actual image needs to be resizedbefore calculation to the size match of the expectedimage with valid values as ("true" and "false").Find (optional)Declares an image-in-image search.2.2.4 Block based identity check:This algorithm is used to partition theimage into quadratic blocks as a selectable size.Each color value of these quadratic blocks isevaluated with the color values average pixelswhich the block contains. If the breadth or height ofthe image is not a multiple block size, then theseblocks are at the right and bottom edge that will becropped and can be weighted accordingly. With theconcluding part the actual blocks are evaluated with
  5. 5. Gopika Mane, Krunank Panchal, Sanchita Sable / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.292-297296 | P a g ethe expected blocks. The final results contain theamount of blocks identical and are divided with thetotal block amount. The algorithm purpose is toevaluate the image which differs for only someparts and still is identical for the remaining parts.Example-The algorithm exemplaryAlgorithm=block; Size=10; Expected=0.95Ravages in the following image:Fig 6 Block based identity check algorithmParametersAlgorithm=blockShould use the Block-based identity check‟algorithmSizeSize of each block is defined with valid valuesbetween 1 and the image size.ExpectedThe minimum match possibility for the check tosucceed is defined with valid values between 0.0and 1.0.Resize (optional)When the actual image needs to be resized beforecalculation to the size match of the expected imageis defined with valid values between ( "true" and"false").Find (optional)Required when an image-in-image search2.2.5 Block based similarity check:This algorithm is used to partition theimage as per the quadratic blocks with a selectablesize. Every value of color blocks is evaluated withthe color average values of the pixels that blockcontains. When the evaluated breadth or height ofthe image is not the multiple with the block size,then blocks contained at the right and bottom edgecan be cropped and may be weighted as per it. Eachexpected block with color value is evaluated withthe actual (original) block. These color values willbe analyzed for similarity percentile. The finalresultant contains all similarity average blocks withtheir weight taken into consideration. Thisalgorithm is used for evaluating the images withsimilar variances color.Example-The algorithm exemplaryAlgorithm=block similarity; Size=5;Expected=0.95Ravages in the following image:Fig 7 Block based similarity check algorithmParametersAlgorithm = block similarityShould use the Block-based similarity checkalgorithmSizeSize of each block needs to be defined betweenvalid values (1 and the image size).ExpectedThe minimal match probability for the check tosucceed should be defined as per the valid valuesbetween (0.0 and 1.0).Resize (optional)When the actual image needs to be resized beforeevaluation to match the size of the expected imageas per the valid values between ( "true" and"false").Find (optional)Should define an image-in-image search.2.2.6 Histogram check:The need to create a histogram checkrequires the image to be first broken down into itsthree build colors red, green and blue. After wardsthe values of the color for each pixel will beanalyzed and partition them into a defined amountof categories (that is referred as buckets when weare talking about histograms). The actual level fillof each bucket can be compared with the expectedlevel. The result of this algorithm will be comparedas the relative frequencies of color categories.Histograms are used for many different scenarios.For example if the need to check tendencies coloror increase brightness. At this situation histogramsmay not be used for evaluating rather than plain-colored images.Example-The algorithm exemplaryAlgorithm=histogram; Buckets=64; Expected=0.95Ravages in the following image:Fig 8 Histogram Check algorithmParametersAlgorithm=histogramImage check should use a Histogram algorithm.BucketsNeed to define how many buckets are requiredbetween valid values as the power of (2 between 2and 256).
  6. 6. Gopika Mane, Krunank Panchal, Sanchita Sable / International Journal of EngineeringResearch and Applications (IJERA) ISSN: 2248-9622 www.ijera.comVol. 3, Issue 3, May-Jun 2013, pp.292-297297 | P a g eExpectedThe minimum match possibility for the check tosucceed should be defined between valid values(0.0 and 1.0).Resize (optional)When an actual image needs to be resized beforeevaluation to match the size of the predictableimage with thevalid values as ("true" and "false").Find (optional)When an image-in-image search is defined.2.3 Cluster Matching Algorithm:A new method for estimatingdisplacements in computer imagery through clustermatching is presented. Without reliance on anyobject model, the algorithm clusters two successiveframes of an image sequence based on position andintensity. After clustering, displacement estimatesare obtained by matching the cluster centersbetween the two frames using cluster features suchas position, intensity, shape and average gray-scaledifference. The performance of the algorithm wascompared to that of a gradient method and a blockmatching method. The cluster matching approachshowed the best performance over a broad range ofmotion, illumination change and objectdeformation. In this study, clustering is used as ameans to group pixels with similar feature„s in eachimage frame; the displacements are estimated bymatching clusters between two frames in the imagesequence. The method should also be applicable tomotion estimation for images containingdeformable objects with varying brightness. First,pixels in each image frame are clustered the twosuccessive frames are treated independently. Eachof the resulting clusters tends to be spatiallyconnected because of the position constraint andcontains pixels of similar gray-scale values becauseof the intensity constraint. Second, clusters arematched between the two consecutive frames. Inaddition to the position and the intensity, the shapeof each cluster and the average gray-scaledifference between two clusters are used asmatching criteria. The Displacement between thetwo matching cluster centers is assigned to everypixel in the cluster. This results in a dense set ofdisplacement estimates.[10]Mathematical module for Cluster MatchingAlgorithmis the called model energy and isdesigned to provide spatiotemporal regularity in themotionis called fittness energy and isdesigned to ensure a certain level of attachment tothe input data, i.e. the observation.3. CONCLUSIONThis paper discuss various pixel matchingalgorithms and Image matching algorithms such asK-Means, Histogram Check and Cluster MatchingAlgorithm. These are required to implementdifferent functionalities of Video SurveillanceSystem Application.ACKNOWLEDGEMENTSWe are thankful to the staff and students ofMarathwada Mitra Mandal‟s Institute of Technology,Lohagaon, Pune College for the support andencouragement from themREFERENCES[1] Barnes, C., Shechtman, E., Finkelstein, A.,Goldman, D.: “PatchMatch: a randomizedcorrespondence algorithm for structuralimage editing.” ACM Transactions onGraphics (Proc. SIGGRAPH) 28 (2009)24[2] Baker, S., Scharstein, D., Lewis, J., Roth,S., Black, M., Szeliski, R.: “A databaseand evaluation methodology for opticalow”. In: Proc. ICCV. Volume 5. (2007)[3] Buades, A., Coll, B., Morel, J.: “A non-local algorithm for image denoising.” In:Proc. CVPR. (2005) II: 60[4] O. Boiman and M. Irani, “Detectingirregularities in im- ages and in video”,ICCV, 2005.[5] S.Agarwal, A. Awan, and D. Roth,“Learning to detect objects in images via asparse, partbased representation.” PAMI,26(11):1475–1490, 2004.[6] J.K. Aggarwal and Q. Cai, “Human motionanalysis: a re- view,” CVIU, Vol. 73, No. 3,pp. 428-440, 1999.[7] J.Beis and D. Lowe, “Shape IndexingUsing Approximate Nearest-NeighborSearch in High- Dimensional Spaces.” InCVPR, pages 1000–1006, 1997.[8] V. Bhaskaran and K. Konstantinides,“Image and Video CompressionStandards: Algorithms andArchitectures,” Kluwer, 1997.[9] O.D.Faugeras,”Three dimensionalComputer Vision,” MIT Press, 1993.[10] L.G Brown,”A Survey of imageregistration techniques,” ACM computingsurveys ,vol.24,no.4,pp.325-376,Dec-1992[11] J.L. Bentley, “Multidimensional binarysearch trees used for associativesearching.” In Communications of theACM, 18(9):509–517,1975