Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 10, No. 2, April 2020, pp. 1337 1345 ISSN: 2088-8708, DOI: 10.11591/ijece.v10i2.pp1337-1345 r 1337 Rob ust f or egr ound modelling to segment and detect multiple mo ving objects in videos Rahul M P atil, Chethan K P , Azra Nasr een, Shobha G Department of Computer Science and Engineering, Rashtree ya V idyalaya Colle ge of Engineering, Bang alore, Karnataka, India Article Inf o Article history: Recei v ed Dec 9, 2018 Re vised May 31, 2019 Accepted Oct 5, 2019 K eyw ords: Background subtraction F ore ground modelling Mean A v eraging Mo ving object detection V ideo analysis ABSTRA CT Last decade has witnessed an e v er increasing number of video surv eillance installa- tions due to the rise of security concerns w orldwide. W ith this comes the need for video analysis for fraud detection, crime in v estig ation, traf fic monitoring to name a fe w . F or an y kind of video anal ysis application, detection of mo ving objects in videos is a fundamental step. In this paper , an ef ficient fore ground modelling method to se gment multiple mo ving objects is implemented. Proposed method signif- icantly reduces noise thereby accurately se gmenting re gion of interest under dynamic conditions while handling occlusion to a lar ge e xtent. Extensi v e performance analysis sho ws that the proposed method w as found to gi v e f ar better results when compared to the de f acto standard as well as relati v ely ne w approache s used for mo ving object detection. Copyright c 2020 Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Rahul M P atil, Department of Computer Science and Engineering, Rashtree ya V idyalaya Colle ge of Engineering, Bang alore-560069, Karnataka, India. Email: patilmrahul06@gmail.com 1. INTR ODUCTION The first step in an y video analytics solution is the se gmentation of mo ving objects. Though this has been studied for se v eral y e ars, there has been lot of concerns when accurately detecting mo ving objects such as background noise, illumination changes, v ariable frame rate in recording videos resulting in lag, shado ws and occlusion to name a fe w . In this paper , we propose an ef fic ient object detection method that addresses issues such as background noise, illumination changes/reflection causing f alse positi v es, o v erlapping or occlusion to lar ge e xtent, e xtracting e xact bounding box or re gion of interest (R OI) using morphological operations and con v e x hull algorithm in post-processing phase. V arious methods ha v e been proposed for back- ground subtraction [1,2], each ha ving its o wn limitation due to man y challenges such as sudden changes in scene, non-static background objects , lag introduced due to v ariable frame rate, changes in appearance of the objects with vie wpoint and dynami c backgrounds such as gush of wind, mo v ement of tree lea v es, shado ws etc. A re vie w of the most rele v ant methods in background subtraction is pro vided in [3], gi ving a good under - standing of the opt imal method to be used for an y background subtraction task. Se gmentation methods using techniques such as background subtraction, Deep Learning etc., play highly pi v otal roles in se v eral applications, ranging from visual observ ation of animals [4,5] to video surv eillance syst ems [6,7]. The y are also e xtremely popular in content based video coding as in [8,9]. Much of the past and on-going research in this field aims at resolving these issues in order to impro v e accurac y of results [10]. Gaura v T akhar et al [11] discusses v arious methods of background subtraction such as basic, statistical as well as the machine learning techniques with the a v erage, best and w orst cases of se v eral J ournal homepage: http://ijece .iaescor e .com/inde x.php/IJECE Evaluation Warning : The document was created with Spire.PDF for Python.
1338 r ISSN: 2088-8708 other dif ferent methods. Proposed system is compared with statistical technique of adapti v e Gaussian mixtures using popular datasets. Non-max suppression technique is discuss ed in [12]. A f aster v ersion of this method helps in the process of mer ging bounding box es if multiple bounding box es are obtained for a single object, which are in close proximity and ha v e similar area sizes. F or se v eral morphological transformations that are used in the proposed method, sound understanding of these are pro vided in [13], most popular being Gaussian mixture model [14]. The state of the art in background subtraction has been proposed by [15], where an adapti v e Gaussian mixture model is used to automatically find the number of Gaussi an components for each pix el. A subsequent method is described in [16], where ef ficienc y of the adapti v e Gaussian mixture model is impro v ed. Arun V ar ghese et al [1] discusses background subtraction being done at the pix el le v el and performance analysis using popular dataset Highw ay from changedetection.net . Performance analysis at the pix el le v el is also discussed in [17]. W e used Pedestrians and Highw ay dataset from baseline cate gory and T urnpik e from the lo w frame rate cate gory of the 2014 CD W da tasets. Frame based performance metrics are discussed in [18,19] such as T rue Positi v es, F alse Positi v es, F alse Ne g ati v es and T rue Ne g ati v es for dif ferent datasets and models respecti v ely . The system proposed in this paper uses techniques such as f ast non-maximum suppression method to increase the ac curac y of detection, con v e x hull method to get better defined blobs of each fore ground object and morphological transformations wit h circular k ernels to get a much smoother outline of the detected fore ground blobs. The model is e xtremely lightweight, v ery f ast and requires no initial training. Proposed model also accounts for changing background by ha ving the background updated by using weighted a v erages of each input frame. All in all, the model is computationall y ef ficient, accurate for major ity of the cases with a small number of limitations that will be discussed later . 2. GA USSIAN MIXTURE MODEL Pix els in the background are modelled with a mixture of K Gaussian distrib utions, the v alue of K being three to v e. The time that a pix el stays in the scene is determined by the weights of the distrib uti ons in the mixture. The most lik ely background colours will be the ones that stay longer as determined by the weights. Impro v ed Gaussian mixture model is more adapti v e than the Gaussian mixture model [15,16], K distrib utions used for modelling is appropriately determined for each pix el in the image. The probability of a pix el ha ving v alue X N at time N is indicated in equation (1): p ( X N ) = K X j =1 w j ( X N ; j ) (1) Wherein w k is weight k th Gaussian component. ( x ; k ) is normal distrib ution of k th component as indicated in equation (2): ( x ; k ) = ( x ; k ; P k ) = 1 (2 ) D 2 j P k j 1 2 e 1 2 ( x k ) T P 1 k ( x k ) (2) In which mean is k and the co v ariance is P k = 2 k I . The K distrib utions are sorted based on the v alue of w k / k and the first B distrib utions are used to create a model of the background of the scene. B is computed as in equation (3): B = arg min b 0 @ b X j =1 w j > T 1 A (3) Where T is the minimum fraction of the background model. In other w ords, it is the mi nimum prior probability that the background is in the scene. (a) GMM adapti v e to v ariable lighting conditions: This method incorporates per pix el Bayesian se gmen- tation into the Gaussian mixture model in order to account for videos recorded in v ariable lighting conditions [20]. (b) Adapti v e v ariable frame rate coding: This method adjusts the frame-rate of the video dynamically and adapti v ely , making use of information from already e xisting video encoders [21]. Int J Elec & Comp Eng, V ol. 10, No. 2, April 2020 : 1337 1345 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1339 (c) Intermittent motion coding: This method in v olv es disabling of motion coding during periods of inacti vity in the video. Thus it records only parts of the video were acti v e fore ground mo v ement is in v olv ed for further processing [22]. All of the methods e xplained abo v e incur considerable o v erhead with re g ard to time or CPU usage. The Gaussian mixture model based methods cannot ef ficiently deal with v ariable frame rates in videos. The v ariable frame rate coding techniques mak e use of video encoder information, the compilation of which in v olv es CPU o v erhead. Also, recording only during periods of acti vity means that the definiti on of acti vity in the scene has to be pre-determined in adv ance, and done so using e xtensi v e statistical analysis. Non-static background objects must be included the background modelled. 3. PR OPOSED SYSTEM Background is modelled by obtaining the background scene without occurrence of an y of the fore ground objects, so that fore ground objects from it can be obtained by background subtraction. Though it looks simple, it is v ery dif ficult and a tedious task as it should not contain an y fore ground objects in it, i.e an y mo v ement suc h as gush of a wind, mo v ement of tree lea v es etc. should be part of the background itself. The background of the scene should be updated as and when the scene changes and must be free from an y kind of noise and must be susceptible to an y kind of illumination changes. 3.1. Running a v erage method A background model has to be constructed initially in order to perform the background subt raction task. Running a v erage is found to be a good method of approximating the background. This method is f aster than Gaussian mixture model and is more consistent than direct frame dif ferencing [23]. Proposed system uses f ast running a v erage method for background modelling as illustrated in equation (Eq. 4): dst ( x; y ) = (1 r ) :dst ( x; y ) + r :sr c ( x; y ) (4) Where dst ( x; y ) is the accumulator image with the same number of channels as input image, sr c ( x; y ) is input image which can ha v e 1 or 3-channels, and r is a weight of the input image. Using continuous frames in a video stream, the weighted a v erage background model can be calculated by choosing an appropriate v alue for r , for that particular sequence. By using a higher v alue of r , we are able to eliminate the fore ground objects that are not persistent in the scene. Also, a suitable v alue of r can be chosen by taking into consideration the amount of data a v ailable for modelling. The process of learning the background is as illustrated in Figure 1. Figure 1. Running A v erage to learn background 3.2. Backgr ound subtraction The V channel of the HSV image is fed as an input to the dif ferencing method, where the absolute dif ference between the V channel of the current frame and the modelled background is obtained. This is done by finding the abs olute dif ference between each pix el element of the modelled background and the V channel Rob ust for e gr ound modelling to se gment and detect multiple ... (Rahul M P atil) Evaluation Warning : The document was created with Spire.PDF for Python.
1340 r ISSN: 2088-8708 of the current frame, which are fed as parameters to the method. The HSV color space is used because it w orks well ag ainst shado ws [24]. The final absolute dif ferenced image is processed to find and dra w the most prominent contours for the detected fore ground objects. Then a thresholding is performed where pix els belo w a certain threshold v alue are assigned a 0 v alue, and the pix els ha ving a v alue greater are assi gned the maximum v alue of 255. This method is kno wn as binary thresholding as sho wn belo w: dst ( x; y )   ( maxV al if ; sr c ( x; y ) > thr esh 0 o ther w ise (5) Here sr c ( x; y ) is a source image pix el, thr esh is the threshold v alue used in binary thresholding and dst ( x; y ) is the result image pix el. maxV al is the v alue that the particular sr c ( x; y ) pix el will obtain if it’ s v alue e xceeds that of the pre-assigned thr esh v alue. The entirety of the steps performed in the proposed method can be e xpressed in a flo w diagram as seen in Figure 2. The Sequence of operation are sho wn in Figure 3. Figure 2. Proposed method to se gment mo ving objects (a) Dif ference Image (b) Gaussian Blur Applied (c) Thresholding, remo ving noisy contours and opening operation (d) Final contours obtained after Con v e x Hull Figure 3. Sequence of operations 3.3. F or egr ound modelling After the threshold frame is determined, we ha v e a binary frame with blobs representing fore ground objects. Morphological transformations such as dilation, erosion and opening are applied to reduce mer ging of contours of dif ferent fore ground objects. Opening operation is used to eliminate portions of the fore ground object that may just e xtend out into the background. It is achie v ed by using the dilation and erosion operation which augments and shrinks a re gion respecti v ely . W e use a structuring element S otherwise kno wn as a k ernel to perform these operations. This operation is used to e xpand the fore ground object’ s obtained contours. The dilation of an image B with S , is gi v en by the belo w equation (6): B S = [ bB S b (6) Int J Elec & Comp Eng, V ol. 10, No. 2, April 2020 : 1337 1345 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1341 Erosion reduces the size of the fore ground object’ s contour and is used to remo v e unw anted e xcess contour elements that may ha v e e xtended into the background. Similar to dilation, the erosion of an image B with structuring element is gi v en belo w: B S = f b j b + s B 8 s S g (7) Opening operation is erosion operation follo wed by dilation operation, which is used to pre v ent mer ging of contours of dif ferent objects and ultimatel y gi v es much better final bounding box es for the fore ground objects, and can be represented mathematically as in equation (8): B S = ( B S ) S (8) These blobs are e xtract ed as contours. Smaller blobs and contours that represent noise and other unw anted detail are eliminated and properti es lik e the edges, centres and areas of the final set of resulting contours are calculated. A con v e x hull of the contours is found to gi v e a definiti v e shape to an y incomplete contours that might ha v e resulted due to similarity of intensity v alue or illumination defects. In order to get whole bounding box es for fore ground objects, there w as a need to mak e the contours of the fore ground objects more wholesome. T o accomplish this, the con v e x hull operation is performed on the cont ours. The contours obtained finally after performing this are used to dra w the bounding box es for the detected fore ground objects. The con v e x hull of a finite set of points S is the set of all con v e x combinations of the points. Each and e v ery point in this set denoted by x i is attrib uted with a weight i . Each and e v ery weight must be non-ne g ati v e and their sum must be equal to unity . These weights are used to obtain a weighted a v erage of all the points in set S . F or v arious choice of coef ficients, a certain con v e x combination is obtained that is a point in the con v e x hull. Therefore, the entire con v e x hull may be obtained by considering all the v arious combinations of weights. It can be e xpressed in a single equation as sho wn belo w in equation (9): C onv ( S ) = j S j X i =1 D i x j j ( r i : i 0) ^ j S j X i =1 i = 1 E (9) The final blobs are returned as contours, and the bounding box es for all these contours are obtained and stored in an array structure. Then re d undant bounding box es that occur inside other lar ger bound- ing box es are eliminated. Finally an iteration of f ast non-max-suppression is emplo yed to mer ge multiple detections for the same object for impro v ed final results. It uses area of the obtained box es in addition t o the o v erlapping percentage of neighbouring box es. Then the final box es that are in the array are dra wn onto the frames. Area of these bounding box es along with their pix els are compared with the bounding box es and the pix els of the ground truth frames in order to estimate and analyse the performance. 4. EXPERIMENT AL SETUP AND RESUL T AN AL YSIS Dataset used for performance e v aluation is CDnet, (Change Detection), consists of 31 videos depicting indoor and outdoor scenes with boats, cars, trucks, and pedestrians that ha v e been captured in dif ferent scenarios and contain a range of challenges. Pedestrians and Highw ay from baseline cate gory and T urnpik e from the lo w frame rate cate gory of the 2014 CD W datasets ha v e been us ed. The v alidation metrics that ha v e been used in the conte xt of comparing the se gmented result with the corresponding ground-truth for that frame in the video sequence are: (a) T rue Negati v e (TN) : Pix els correctly classified as the background (b) T rue P ositi v e (TP) : Pix els correctly classified as the fore ground (c) F alse P ositi v e (FP) : Pix els wrongly classified as the fore ground (d) F alse Negati v e (FN) : Pix els wrongly classified as the background V arious performance metrics that ha v e been used are as sho wn from equation (10) to (17) belo w: P r ecision ( P ) = T P F P + T P (10) Rob ust for e gr ound modelling to se gment and detect multiple ... (Rahul M P atil) Evaluation Warning : The document was created with Spire.PDF for Python.
1342 r ISSN: 2088-8708 R ecal l ( R ) = T P F N + T P (11) S pecif icity = T N F P + T N (12) F al se N eg ativ e R ate = F N F N + T P (13) F al se P ositiv e R ate = F P F P + T N (14) P W C = F P + F N T N + T P + F P + F N 100 (15) F M easur e = 2 R P R + P (16) Accur acy = T N + T P T N + T P + F N + F P (17) T able 1 sho ws the performance comparison of the proposed system, ag ainst the impro v ed adapti v e Gaussian mixture model [15] on three datasets, namely highw ay , turnpik e and pedestrians. T able 1. Performance e v aluation of proposed system with impro v ed adapti v e Gaussian mixture model and Hybrid model Datasets Highway P edestrians T ur npik e Model Pr oposed Zi vk o vic[15] Hybrid Pr oposed Zi vk o vic Hybrid Pr oposed Zi vk o vic Recall 0.7387 0.9619 0.9152 0.6594 0.9860 0.7290 0.9259 0.9649 Specificity 0.9982 0.9272 0.9314 0.9988 0.9613 0.9921 0.9868 0.9695 FPR 0.0137 0.5682 0.5391 0.0216 0.6804 0.1384 0.0724 0.1678 FNR 0.0334 0.0049 0.0118 0.0194 0.0008 0.0154 0.0134 0.0064 PWC 3.1237 6.8897 7.0895 1.9496 3.7379 2.2092 2.2525 3.1197 Pr ecision 0.9817 0.6286 0.6293 0.9682 0.5917 0.8404 0.9275 0.8519 F-Measur e 0.8430 0.7603 0.7453 0.7845 0.7396 0.781 0.9267 0.9049 Accuracy 0.9688 0.9311 0.9288 0.9792 0.6258 0.9767 0.9775 0.9688 As indicated in T able 1 it w as found that proposed method w as found to be ef fecti v e and yielded better accurac y of 96.88% and precision of 98.17%. Also, i t has a v ery lo w f alse positi v e rate and f alse ne g ati v e rate for detecting m o vi n g objects in videos, when compared to the de f acto standard of the impro v ed adapti v e Gaussi an mixture model on the highw ay dataset from change detection net. The snapshots obtained with proposed system, and the adapti v e Gaussian mixture model for three datasets, as sho wn in Figures 4, 5 and 6. T able 1 also sho ws comparison of the proposed system with another e xisting method, namely the multi-modal h ybrid approach of adapti v e Gaussian mixture model and mean a v eraging. The h ybrid model used for comparison can model and track mo ving objects in a video and it w orks as follo ws. In order to smoothen the e xtracted frames, a sequence of smoothing filters are applied, these being Gaussian blur and median blur , respecti v ely . The approach tak en to reduce noise uses the morphological operations erosion and dilation. Mean a v eraging is used for background modelling and frame dif ferencing along with the adapti v e Gaussian mixture model is used to obtain fore ground masks. Contours are found from the fore ground masks on which con v e x hull is applied to get the final object blobs. Proposed h ybrid model is able to detect and track mo ving objects in videos in real time and is tested for man y outdoor scenes, and snapshots of the obtained results follo w the conclusion section. No comparison has been made for the T urnpik e dataset for the h ybrid model, as it has not been designed for lo w frame rate videos, and therefore it has not been included in the table. As e vident from the T able 1, the proposed system is able to perform well when compared to the h ybrid method as well. Ef fec ti v ely reduces noise and is able to se gment e xact R OI of mo ving objects. This is achie v ed by Gauss ian blur and remo v al of small contours leading to noise, and by applying opening morphological operations. This isolates contours of dif ferent bounding box es, e v en if the distance between the objects is small, thereby handling occlusion to an e xtent. Int J Elec & Comp Eng, V ol. 10, No. 2, April 2020 : 1337 1345 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1343 (a) Original Input (b) Ground T ruth (c) Contours obtained by MoG2 (d) Contours obtained by Proposed Model (e) Bounding box es from Ground T ruth (f) Bounding box es from MoG2 (g) Bounding box es from Proposed Model Figure 4. Results for T urnpik e dataset (a) Original Input (b) Ground T ruth (c) Contours obtained by MoG2 (d) Contours obtained by Proposed Model (e) Bounding box es from Ground T ruth (f) Bounding box es from MoG2 (g) Bounding box es from Proposed Model Figure 5. Results for T urnpik e dataset Rob ust for e gr ound modelling to se gment and detect multiple ... (Rahul M P atil) Evaluation Warning : The document was created with Spire.PDF for Python.
1344 r ISSN: 2088-8708 (a) Original Input (b) Ground T ruth (c) Contours obtained by MoG2 (d) Contours obtained by Proposed Model (e) Bounding box es from Ground T ruth (f) Bounding box es from MoG2 (g) Bounding box es from Proposed Model Figure 6. Results for T urnpik e dataset The Figures 4, 5 and 6 sho w a comparison of the w orking of our proposed model ag ainst Impro v ed Adapti v e Gaussian mixture model and the Hybrid model. Each figure consists of a set of 7 sub-figures each, which summarize the results obtained on the dif ferent datasets that ha v e been used. The first sub-figure, is the input frame from the original dataset, just as is the follo wing ground truth sub-figure. The follo wing tw o sub- figures are the blobs that are obtained by the impro v ed adapti v e Gaussian mixture model and our o wn method respecti v ely . The follo wing three figures are as their captions suggest. Essentially , the y are bounding box es that ha v e been obtained for the corresponding blobs, and dra wn onto the original input frame. 5. CONCLUSION The proposed system w as found to be an ef fecti v e approach in capturing small and lar ge mo v ements in the mo ving objects and e xtracts well defined fore ground objects. Exact re gion of interest were e xtracted and it yielded better accurac y when compared to state of art de v elopment method such as mixture of Gaussians and relati v ely ne w h ybrid approach of mean a v eraging and mixture of Gaussians method when it comes to issues such as noise and much better contours when considering indi vidual and multiple objects. An y noise due to flick ering of frames or noises added to the camera feed are ef fecti v ely remo v ed from being included in the fore ground. The mer ging of fore ground objects that might tak e place due to occlusion of multiple fore ground objects has been a v oided to a maximum e xtent using morphological transformations. The proposed model is a light weight model which can perform background subtraction in real time on machines with v ery basic processing po wer . Future enhancement can be shado w detection and better splitti ng of contours of objects that are totally occluded. REFERENCES [1] Arun V ar ghese, Sreelekha G, ”Background Subtraction for V ehicle Detection, Pr oceedings of Global Confer ence on Communication T ec hnolo gies 2015 (GCCT 2015) , pp. 380-382, 2015. [2] Azra Nasreen, Kaushik Ro y , K unal Ro y , Shobha G, ”K e y Frame Extraction and F ore ground Modelling Using K-Means Clustering, 7th International Confer ence on Computational Intell ig ence Communication Systems and Networks (CICSyN) , pp. 141-145, 2015. [3] Massimo Piccardi, ”Background Subtraction T echniques: A Re vie w , IEEE International J ournal on Sys- tems, Man and Cybernetics , V ol. 2, (5), pp. 05-25, 2004. [4] T . K o, S. Soatto, D. Estrin, ”Background Subtraction on Distrib utions, Eur opean Confer ence on Computer V ision (ECCV 2008) , pp. 222-230, October 2008. Int J Elec & Comp Eng, V ol. 10, No. 2, April 2020 : 1337 1345 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1345 [5] M. Himmelsbach, U. Knauer , F . W inkler , F . Zautk e, K. Bienefeld, B. Mef fert, ”Application of an Adapti v e Background Model for Monitoring Hone ybees, VIIP 2005 , 2005. [6] Q. Ling, J. Y an, F . Li, Y . Zhang, ”A Background Modelling and F ore ground Se gmentation Approach Based on the Feedback of Mo ving Objects in T raf fic Surv eillance Systems, Neur ocomputing , 2014. [7] Rahul M P atil, N R V inay , Rohith Y , Ram Srini v as, Pratiba D, ”IoT Enabled V ideo Surv eillance System us- ing Raspberry Pi, 2nd Confer ence on Computational Systems and Infor ma t ion T ec hnolo gy for Sustainable Solutions (CSITSS 2017) , December 2017. [8] S. Chakraborty , M. P aul, M. Murshed, M. Ali, ”An Ef ficient V ideo Coding T echnique Using a No v el Non-parametric Background Model, IEEE International Confer ence on Multimedia and Expo W orkshops (ICMEW 2014) , pp. 1-6, July 2014. [9] X. Zhang, Y . T ian, T . Huang, W . Gao, ”Lo w-comple xity and High-ef ficienc y Background modelling for Surv eillance V ideo Coding, IEEE International Confer ence on V isual Communication and Ima g e Pr o- cessing (VCIP 2012) , San Jose, USA, No v ember 2012. [10] T . Bouwmans, ”T raditional and Recent Approaches in Background modelling for F ore ground Detection: An Ov ervie w , Computer Science Re vie w , 2014. [11] Goura v T akhar , Chandra Prakash, Namita Mittal, Rajesh K umar , ”Comparati v e Analysis of Background Subtraction T echniques and Applications, IEEE International Confer ence on Recent Advances and Inno- vations in Engineering (ICRAIE-2016) , pp. 1-8, 2016. [12] Pedro F . Felzenszw alb, Ross B. Girshick, Da vid McAllester , De v a Ramanan, ”Object Detection with Discriminati v ely T rained P art Based Models, IEEE T r ansactions on P attern Analysis and Mac hine Intel- lig ence , V ol. 32, (9), pp.1627-1645, 2010. [13] Linda Shapiro et al., editors, ”Computer V ision, Illustr ated, Oxfor d UP , Pr entice Hall , 2001. [14] Zezhi Chen, T im Ellis, ”A Self-Adapti v e Gaussian Mixture Model, International J ournal of Else vier Computer V ision and Ima g e Under standing , V ol. 122, (3), pp. 35-46, 2014. [15] Zi vk o vic Z, Impro v ed Adapti v e Gaussian Mixture Model for Background Subtraction, Pr oceedings of International Confer ence on P attern Reco gnition (ICPR) , Mosco w , pp. 28-31, 2004. [16] Zi vk o vic Z, ”Ef ficient Adapti v e Density Estimation per Image Pix el for the T ask of Background Subtrac- tion and P attern Recognition Letters, International J ournal on P attern Reco gnition (IJPR) , V ol. 27, (7), pp. 773-780, 2006. [17] N. Go yette, P .-M. Jodoin, F . Porikli, J. K onrad, and P . Ishw ar , http://changedetection.net , Pr oc. IEEE W orkshop on Chang e Detection (CD W -2012) at CVPR-2012 , Pro vidence, RI, June 2012. [18] F aisal Bashir , F atih Porikli, ”Performance Ev aluation of Object Detection and T racking Systems, TR2006-041, Mitsubishi Electric Resear c h Labor atories , 2016. [19] Haixia W ang, Li Shi, ”F ore ground Model for Background Subtraction wit h Blind Updating, IEEE Inter - national Confer ence on Signal and Ima g e Pr ocessing , pp. 74-78, 2016. [20] A B Godbehere, Matsuka w a A, Goldber g K, ”V isual T racking of Human V isitors Under V ariable Lighting Conditions for a Responsi v e Audio Art Installation, American Contr ol Confer ence (A CC) , pp. 4305-4312, June 2012. [21] Y u Y uan, Feng D, Y uzhuo Zhong, ”F ast Adapti v e V ariable Frame Rate Coding, IEEE V ehicular T ec hnol- o gy Confer ence , V ol. 5, pp. 2734-2738, May 2004. [22] Guarangnella C, Di Sciasco E, ”V ariable Frame Rate for V ery Lo w Bit Rate V ideo Coding, 10th Mediter - r anean Electr otec hnical Confer ence , V ol. 2, pp. 503-506, 2000. [23] Zheng Y i, F an Liangzhong, ”Mo ving Object Detection Based on Running A v erage Background and T em- poral Dif ference, Pr oceedings of International Confer ence on Intellig ent Systems and Knowledg e Engi- neering (ISKE) , T aiw an, pp. 270-272, 2010. [24] V inod M, Sra v anthi T , Brahma Reddy , ”An Adapti v e Algorithm for Object T racking and Counting, International J ournal of Engineering and Inno vative T ec hnolo gy (IJEIT) , V ol. 2, (4), pp. 560-585, 2012. Rob ust for e gr ound modelling to se gment and detect multiple ... (Rahul M P atil) Evaluation Warning : The document was created with Spire.PDF for Python.