Appearance Based Tracking with Background Subtraction

Transcription

1 The 8th International Conference on Compter Science & Edcation (ICCSE 213) April 26-28, 213. Colombo, Sri Lanka SD1.4 Appearance Based Tracking with Backgrond Sbtraction Dileepa Joseph Jayamanne Electronic and Telecommnication Engineering University of Moratwa-Sri Lanka Jayath Samarawickrama Electronic and Telecommnication Engineering University of Moratwa-Sri Lanka Ranga Rodrigo Electronic and Telecommnication Engineering University of Moratwa-Sri Lanka Abstract-Groping the detected featre points traditionally reqires the storage of long corner tracks. The traditional method does not permit to arrive at a decision to clster the featre points based on a frame by frame basis. This paper presents a method to grop the featre points directly into objects sing the most recent 2 frames. The detected corner featres are validated and clstered based on two approaches. When objects move in isolation, an EM algorithm is sed to clster and every object is detected and tracked. When objects move nder partial occlsion, the corner featres are clstered based on an agglomerative hierarchical clstering approach. A probabilistic framework has also been applied to determine the object level membership of the candidate corner featres. A novel foregrond estimation algorithm with an accracy of 98% based on color information, backgrond sbtraction reslt and detected corner featres is also presented. I. INTRODUCTION Monitoring vehiclar traffic sing srveillance cameras still need manal intervention. There have been work ndertaken to detect, localize, and classify vehicles and to analyze vehicle behavior like estimation of average speed, trajectory, flow rate and density sing the video footage obtained throgh cameras [1]-[4]. The prpose of developing sch atomatic traffic srveillance systems is to track vehicles, monitor the vehicle traffic and extract traffic parameters. This system will enable to redce manal intervention in monitoring traffic and identify reglar road sers and traffic violators. Traffic srveillance systems rely on accrate object detection and tracking. A. Related Work Extracting and tracking individal corner featres or interest points and groping them based on their trajectories is a method sed in object tracking [1]. Althogh the segmentation of occlded objects is easy to perform, tracking the same corner featre for a long period of time is challenging. Frthermore a set of long corner trajectories will have to be processed and kept, reslting in heavy se of memory. The corner featres are tracked from point of entry to exit and groped directly into objects after obtaining long trajectories of the corner tracks sing the proximity and motion history [1]. Points that rigidly move forward are groped allowing occlsion to be handled at the cost of comptational memory. This is a typical example of a single level clstering and featre groping approach. Kim [2] considers a dynamic mlti-level featre groping approach to obtain refined trajectories in real time in contrast to Coifman et al. [1]. Emerging featre points are initially groped into small clsters sing a Normalized-ct algorithm [5] and frther a variation of Expectation Maximization algorithm is applied to serve two prposes namely; i) contine clstering the same clsters previosly detected by N-ct in the next set of frames and ii) grop the clsters that are detected to achieve object level groping. Backgrond sbtraction provides the base for most of the object tracking algorithms. Initially, it extracts a backgrond hypothesis from a seqence of frames. The difference of the backgrond hypothesis and the crrent frame separates the foregrond. Althogh the comptational time it reqires is relatively small, it is nable to deal with occlsions, shadows, and sdden illmination changes. Kim [2] has combined the backgrond sbtraction and the featre tracking and groping algorithms to prodce high qality object trajectories from fragmented featre tracks. Kim's agmentation to the backgrond sbtraction algorithm ses a low-level featre tracking as a ce to validate the estimated foregrond region; however, Kim has estimated the silhoettes based on conventional morphological operations. The main drawback of this estimation is considering the excess regions otside the bondaries of the objects as the foregrond. This reslts in classifying sch actal backgrond pixels as a part of the foregrond. Tracking applications can be condcted sing a fixed featre space. However, Collins et al. [3] have proposed a method for evalating on-line, adaptive selection of appropriate featre spaces for tracking and for adjsting the set of featres sed to improve tracking performance. They have claimed that the featres that best discriminate the object and the backgrond are best for tracking the object. Althogh a wide range of featres can be sed for tracking like color, textre, shape, and motion, they have only considered a linear combinations of camera R, G, B pixel vales to compose the set of seed candidate featres. Selecting the right featres plays an important role not only in tracking bt also in detection. In general, the most desirable property of a visal featre is its niqeness in order for the /13/$ IEEE 643

2 SD1.4 objects to be easily distingished from featre space [6]. Bch et al. [4] present another method, 3DHOG, for detection and classification of road sers in rban scenes. This system works even if the appearance of vehicles varies sbstantially with the viewing angle. This is an extension to HOG featre extraction [7] by applying 3D spatial modeling to operate on still images. This overcomes the reliability limitations of motion silhoettes. This is an example of sing a complete different featre space to detect the objects. In this paper we directly grop detected corner featres into objects withot waiting till long corner trajectories are available to clster. Or method takes into consideration both ideas of Kim [2] and Malik et al. [I]. Or focs is to se a single clstering algorithm to achieve object level clstering that cold be applied on a frame by frame basis sing Kim's work as a base. We also propose a method to incorporate color information to threshold the backgrond sbtraction reslt to accrately estimate the foregrond pixels. Althogh we se conventional morphological operations to preserve the shape of the objects, we apply dilation with a small strctring element on or backgrond sbtraction reslt. This redces estimating a larger excess region arond the bondary of the actal object as a part of the foregrond. In or work, we propose methods to incorporate color information together with conventional morphological operations on backgrond sbtraction reslt, to preserve the shapes of silhoettes that correspond to different sized objects and arrive at a better foregrond estimate, to achieve single-level clstering that directly correspond to objects withot sing long corner tracks, to validate clster membership of corner featres based on a probabilistic framework sing Bayesian reasoning. II. METHODOLOGY The blobs that are detected sing backgrond sbtraction are validated throgh the KLT point tracks. These tracks are intern assigned to clsters sing several mechanisms. We have applied an EM algorithm to grop the detected corner tracks for instances where vehicles move in isolation and an agglomerative hierarchical clstering algorithm for instances where vehicles move nder partial occlsion. The otline of the proposed methodology is as follows: Modeling the backgrond and pdating it every 15 frames. 2) Estimating the foregrond region: Generating object blobs sing color information and conventional morphological operations. 3) Detecting corner featres sing the KLT tracker [8]. 4) Validating corner featres and validating the estimated foregrond region. 5) Applying a single level hierarchical clstering algorithm when vehicles move nder partial occlsion. 6) Applying single level clstering based on an EM algorithm for vehicles moving in isolation. 7) Applying Bayesian reasoning to determine the membership of a corner featre for each clster. 8) Validating the clstered featre points based on their membership. 9) Tracking clsters and obtaining the trajectories of the corresponding objects. A. The Backgrond Model Or approach is mainly based on Kim's work [2]. We have implemented the sggested backgrond model with a modification to the backgrond sbtraction algorithm when estimating the actal foregrond region. We pdate the backgrond every 15 frames and the frame rate of the videos considered is 3 fps. The backgrond model sed is as follows [2]: Bt+1 Ic(Bt) = Bt+l Ic((l - a)bt = + ant) when lvlt = 1 when lvlt = where Bt represents the backgrond model at time t, Bt+1 is the next backgrond pdate, Nt is the temporal median of the recent 15 frames, lvlt is the binary moving object hypothesis mask, a is a vale in [, I] and Ic(.) is an illminationcorrection fnction which is applied to each of the R, C, B vales as follows: where kr, kg and kb are determined by voting on ReIR, CeIC, and Bcf B over all the pixels in the images and (Re, Ce, Be) are the pixel vales of the crrent frame. lvlt is a state that each pixel cold occpy. lvit = 1 indicates that the considered pixel is estimated to be a foregrond pixel and if lvlt = the considered pixel is estimated to be a non foregrond pixel. lvlt is generated from the resltant difference image obtained from sbtracting the backgrond from the crrent frame. How we compte lvlt based on color information and morphological operations will be introdced in the next sbsection. The comptation of kr, kg and kb for each R, C, B vales of a frame sed to pdate the backgrond in or implementation is as follows: Average of the mid 5% of the crrent frame Average of the mid 5% of the previos backgrond J) Silhoette generation based on color to preserve shape of the foregrond: Converting the conventional backgrond sbtraction reslt to its binary format by sing a threshold wold reslt in a loss of regions that correspond to the actal foregrond. Even if the fragmented pieces that appear cold be dilated to determine the bondary and be filled in to obtain the object blobs/silhoettes, it is difficlt to define a single strctring element that is capable of dilating and preserving the shapes of different sizes of objects. Therefore we se color together with the conventional morphological operations to preserve the shapes of silhoettes that correspond to different sizes of objects. The proposed mechanism sed to determine the vale of NIt and to improve generating silhoettes is (I) (2) (3) (4) 644

3 SD1.4 where x represents R, G, B color vales of each pixel, Ix is the crrent frame, Bt is the recent backgrond pdate and PCx is the percentage color change compared to the recent backgrond pdate. For each pixel when we obtain the percentage color change compared to the recent backgrond pdate, we select the maximm of the three percentages and apply a threshold to determine whether the pixel belongs to the estimated foregrond or non foregrond region. Then we dilate the resltant binary image sing small strctring elements. Next we fill holes, remove small regions, and apply connected component analysis to label the regions and obtain the blobs. In this manner we validate each blob by detecting corner featres [8], [9]. If we do not detect corner featres in a blob, it is considered as a false foregrond region and NIt is set to for all the pixels within the region otherwise NIt is set to 1 for all the pixels within the validated blobs considered as the final estimated foregrond region. B. Single Level Clstering Groping featre points into objects either cold be direct [1] or it cold be arriving at an intermediate stage of groping before clstering into objects [2]. Or approach attempts to grop the featre points directly into object level clsters immediately after being detected withot monitoring how closely a grop of points appear and rigidly move forward. Or method mainly consists of two algorithms that cold be applied based on the scenario. If all the objects move in isolation, object level clstering cold be directly achieved by applying a two dimensional EM Algorithm [1] on the (X, Y) position coordinates of the featre points. EM reqires not only the nmber of initialization points, bt also their corresponding vales close to the expected positions where clsters need to appear so that the desired clstering cold be achieved. Ths we have sed blob centers as the initialization points since each object corresponds to a single blob when objects move in isolation. When vehicles are partially occlded, blobs corresponding to actal objects are merged and therefore this approach fails in this scenario particlarly becase of not having the correct nmber of initialization points and their initial vales. In order to achieve object level clstering nder partial occlsion, the second algorithm we have sed is an agglomerative hierarchical clstering approach. Each featre point is considered as a different clster and pairs of clsters are merged when moving p the hierarchy [11]. We have considered the x and y position coordinates, speed and the trajectory of each of the featre points when generating the featre matrix. Every instance of the featre matrix has been sed to generate the matrix that encodes a tree of hierarchical clsters. In both clstering algorithms, for a given frame once the featre points are detected, the decision to clster will be made after the featre points have been tracked for next 2 frames. For each frame, featre points will be re-detected. Within the next 2 frames the speed of a featre point is compted sing each of the two recent consective frames. Ths the speed of a featre point at a given instance is considered as the median speed of these set of speeds. Bt each point's trajectory is obtained considering only its next appearance. In hierarchical clstering, the absolte gradient is sed as the featre "trajectory" assming each featre point's path follows a straight-line. J) Corner Featre Detection and Validation: In or work, we have sed corner featres as or featre points. The corner featres are detected and tracked sing the KLT tracker [8], [9]. Or work is based on Birchfield's KLT implementation [8] and the detected corner featres are validated sing the criteria mentioned in [2]. If a detected corner featre cold be tracked thrice and if it does not have a match in the corresponding backgrond image, sch a point is termed as a valid featre point. Sch validated corner points are then sed to validate object blobs that reslts in the removal of false foregrond regions facilitating a more accrate estimation of Mt. 2) Membership of corner featres: Once the detected corner featres are clstered, the clster membership of each featre point is determined based on a probabilistic framework. As in [2], for each clstered featre point, given the parameters, a Bayesian reasoning is applied to compte the posterior probability of the featre point to find ot whether it belongs to the clster or not. The extracted parameters are the ratio of being in the same blob: r the proximity: p the history of motion: m as in Kim's work [2]. The comptation of the posterior probability of the clster membership of a featre point depends on the generated prior distribtions. For each featre point, the prior probability of being a member or not is.5 and ths it indicates the two possible states. Therefore for each parameter, we estimate two individal probability distribtions. One distribtion to represent a featre point being a member and the other to represent a featre point not being a member. We sed a semi-spervised procedre to extract these parameters to generate the probability distribtions. The parameter r-ratio of being in the same blob- indicates that within the next 2 frames and ot of 2, the nmber of times a particlar featre point appearing on the actal object. For each featre point, r is actally the ratio of being in the actal object. When generating r-member distribtion- all sch vales of the featre points for a set of frames are considered. In order to obtain rv r-not a member distribtion- as in before, each featre point is tracked for 2 times and ot of 2, the nmber of times a featre point appearing otside the object is considered. The Fig. 1 shows the histograms obtained after normalizing r to be in [,1] with 2 bins. The parameter p---the proximity- refers to the minimm distance from the ellipse bondary to each clstered featre point. In order to extract this parameter, ellipses are drawn arond the candidate clsters and the minimm distance from 645

4 .... SD1.4 2, r-member r-not a member I I 3 I 2, m-member m-not a member 2, 1,5-1, ,5 1, 5 1,5 1, Normalized r Normalized r Fig. 1: left-probability distribtion obtained for r-ratio of being in the same blob- for a featre point appearing on the actal object, right-probability distribtion obtained for rv r-ratio of not being in the same blob- for a featre point appearing otside the actal object. the ellipse bondary for each of these points is obtained. The distance of a point that lies inside the ellipse is sed to generate the p---member distribtion- and the distance of a point that lies otside the ellipse is sed to generate the rv p-not a member distribtion-. Fig. 2 shows the the histograms obtained after normalizing p to be in [,1] with 2 bins. p-member p-not a member Normalized p Normalized p Fig. 2: left-probability distribtion obtained for p---the proximity- for distance of a featre point that lies inside the ellipse bondary, right-probability distribtion obtained for rv p---non proximity- for distance of a featre point that lies otside the ellipse bondary. The parameter m-the history of motion- refers to the speed of a featre point. We compte the m-member distribtion by tracking each featre point within next 2 frames and obtaining the nmber of speeds of the featre point that lie close to the median vale ot of 2. In order to generate rv m-not a member distribtion-, nmber of speeds of each featre point that do not lie close to the median ot of 2 have been sed. Fig. 3 shows the the histograms obtained after normalizing m to be in [,1] with 2 bins. For each appearance of the featre point, the speed is compted sing two recent consective frames. How similar a particlar speed compared to its median is compted as Normalized m Normalized m Fig. 3: left-probability distribtion obtained for m-the history of motion- for speeds of the featre point that lie close to the median vale, right-probability distribtion for rv m for speeds of the featre point that lie away from the median vale. follows: M = abs(sj - 5) Smax - Smin where (5)!vI is the measre that determines how similar the detected speed to the median speed, j is the nmber of times the speed cold be compted within the next 2 frames, 5 is the median speed, Smax and Smin refer to the maximm and minimm speeds respectively. Therefore the posterior probability of the clster membership of a candidate featre point is compted separately considering the extracted parameters individally as follows: P(xlmember) P(mem b erlx ) = P(xlmember) + P(xl rv member) where x r, p, m. As in Kim's work [2] we have also assmed conditional independence of these extracted parameters. = Therefore the above 3-eqations can be expressed as follows: P(memberlr,p, m) = P(r,p, mlmember) P(r,p, mlmember) + P(r,p, ml rv member) where P(r,p, mlstate) P(rlstate) = x P(pistate) x P(mlstate) C. Clster tracking (6) (7) where state = member / rv member (8) In order to track the same clster continosly, we se both the Eclidean distance between the previos and the crrent positions of the clster centers and the trajectory. In every frame once the featre points are tracked in next frame, it is assmed that the featre points to follow an eqation of a straight line and the gradients of all the featre points are obtained. The likely gradient of the straight line the clster center may follow is estimated to be the median of the obtained set of gradients. For each frame, featre points are detected 646

5 ... SD1.4 and tracked in the next 2 frames to extract the necessary parameters to validate before clstering in the crrent frame. Ths the median speed of all the candidate featre points of a clster is already compted. Therefore the the expected center of the clster in the next frame can be estimated sing the frame rate, crrent speed and the gradient of the clster center. Then by thresholding, the same clster is identified. When compting the speeds of the featre points of an object, we can observe a range of different speeds within it. For instance, the speeds of the featre points appearing on the front end of the vehicle may be dissimilar to the speeds of the featre points appearing on the back-end of the same vehicle. This is mainly de to the perspective effect of the camera. Therefore to achieve robstness, the speed of the clster center is considered to be the median of the speeds of the candidate featre points. III. RESULTS AND ANALYSIS We track vehicles nder two different scenarios, namely, when i) vehicles move in isolation, ii) vehicles move nder partial occlsion. Using EM algorithm and an agglomerative hierarchical clstering algorithm the vehicle tracking reslts were obtained. The video footage sed for the experiment were Kim's video clip [2], specific regions of interest of VIRAT video clips [12] and freshly obtained local footage. We have frther improved Kim's foregrond estimation techniqe by incorporating color information together with the conventional morphological operations. We have also introdced a single level agglomerative hierarchical clstering approach to directly clster corner featres sing the most recent 2 frames. The resltant object blobs obtained after incorporating color information together with the conventional morphological operations and after validating the foregrond with the corner "CH' featre points are shown in the Fig '\:!.. :.. gt. -. '...,.,!rf-.. - '1,' : - :..!:l. Ir'. '. I'I. - -.,. -. :-... L_.'A ': ' '!=l,->;1...!!!j.l: _,, _. I. -::. - / I. ' ;;.t ; t,.... Fig. 4: Binary images represent the estimated foregrond obtained by the improved foregrond estimation techniqe indicating an accrate foregrond estimation compared to the existing work. In order to qantitatively determine the accracy of the algorithm, the state of each pixel of a given frame is considered as a two-class prediction problem in which the otcomes are labeled as positive (P) for a pixel in the actal foregrond and negative (N) for a pixel in the actal backgrond. We have applied the foregrond estimation algorithm on a set of completely different images. The poplation is considered as the total nmber of pixels of the set of images selected. The obtained confsion matrix is given below. TABLE I: The confsion matrix obtained by considering the state of the pixel being foregrond or backgrond-the total nmber of pixels of the set of images considered is Actal Class Foregrond (P) I Backgrond (N) I Predicted Foregrond (TP) I (FP) I Predicted Backgrond 6665-(FN) I (TN) The foregrond estimation algorithm estimates the foregrond with a accracy of 98% according to the confsion matrix. This indicates that when the video camera is fixed, given a frame, the developed color based foregrond estimation techniqe is cable of estimating the actal foregrond region with a accracy of 98%. Or method assmes the position of the camera to be fixed in order to estimate the foregrond region. Therefore when there is a slight movement of the camera, certain false foregrond regions are generated. 2:l ;... ;> p., e c.6 - :B rfj -= r:/j.4.2 ROC crve ROC crve Sensitivity vs. I-Specificity I-Specificity (FP rate) Fig. 5: When percentage color change compared to the previos backgrond is thresholded to 3% the classification ability of the foregrond estimation techniqe increases. Ths a better foregrond estimation cold be arrived at by applying simple threshold on the backgrond sbtraction reslt. In foregrond estimation, when the threshold applied on the obtained maximm percentage color change compared to the recent backgrond pdate is varied, the classification ability of the foregrond estimation algorithm cold be evalated sing a ROC crve (Fig. 5). According to the ROC crve, closer a point to the top left corner, higher the correctness of the classification reslts generated by its threshold vale. Therefore we have sed 3% as the threshold vale to decide whether the pixel belongs to the foregrond or backgrond region. The reslts obtained by applying EM algorithm to clster the validated clster points when vehicles move in isolation is shown in Fig. 6. In order to achieve the clstering of featre points to separately identify an object, it is important to initialize the EM algorithm with a point close to the desired 647

6 SD1.4 clster. Therefore when vehicles move in isolation, we se blob centers to initialize the EM algorithm. In Fig. 6, the validated corner featres and the obtained trajectories are shown. Fig. 6: Trajectories obtained by the EM algorithm for a specific region of interest selected in a VIRAT video clip when vehicles move in isolation. Integer part of the displayed nmber represents the frame at which the vehicle is detected and the non integer part of the nmber indicates the assigned clster nmber when it is detected. The obtained tracking reslt is good and is best viewed in color. Althogh agglomerative hierarchical clstering considers each featre point to be a different clster and pairs of clsters are merged along the hierarchy, it still reqires a certain ctoff vale to determine for what extent the clsters mst be merged. In this implementation the nmber of blobs are sed as the nmber of objects in a frame. If the size of the blob is large, it is eroded and the blob cont is sed to identify the nmber of objects per frame. In order to validate the nmber of blobs, another criteria has been applied as follows: When fresh clsters are formed, based on the trajectory and the speed of the clster center, the likely position of the same clster center appearing in the next frame can be compted. Therefore when new featre points are detected on the crrent frame, a search is performed to find ot whether the detected and validated featre points cold be assigned to the same clster. Ths it allows previosly detected clsters/nmber of objects to be carried forward to the next frame ntil the object disappears from the field of view of the camera. The reslts obtained throgh agglomerative hierarchical are shown below. Fig. 7: Trajectories obtained after applying hierarchical clstering to a specific region of interest in Kim's video. The obtained tracking reslt is satisfactory and is best viewed in color. In Fig. 7, the two vehicles initially appeared being partially occlded. Bt generating the featre matrix sing the x and y position coordinates, speed and the trajectory has allowed the agglomerative hierarchical clstering approach to grop the featre points directly in to objects as reqired. In the comptation of probabilities in Fig. 1 for each ratio -T and rv T, the cont is an integer ranging from [1-2]. Bt each validated featre point is tracked thrice and therefore when the ratios are obtained and normalized to [1] range and 2 bins are sed, no vales are observed in certain normalized T vales corresponding to initial conts. In Fig. 2 the probabilities for each parameter -p and rv p are compted as follows: The vales of minimm distance from the ellipse bondary are continos and therefore once the minimm distance is obtained, it is normalized to a vale that lies between [1]. The probability is assigned by checking to which binned range the normalized vale belongs to and retrieving its corresponding probability vale. We have applied or tracking algorithms for several video clips and several selected regions of interest of VIRAT video clips [12]. All the vehicles that move in isolation are detected and tracked. Bt at certain times when vehicles move nder partial occlsion, the hierarchical clstering algorithm failed simply becase of not correctly estimating the nmber of objects per frame to clster. In sch circmstances, when the nmber of objects to clster are manally adjsted, most of the time, the desired clstering is achieved. IV. CONCLUSION AND FUTURE WORK We have directly made an attempt to achieve object level clstering reslting in object tracking. For a given frame, clster membership of the featre points are compted based on a probabilistic framework. We have tracked each featre point for next 2 frames to extract parameters. Then, we have assigned the probability of being a member by sing the generated probability distribtions. The implemented backgrond model reqires the backgrond sbtraction reslt and the detected and validated corner featre reslt to fnction. The color information of a pixel and the conventional morphological operations have been sed to preserve the shape of the estimated silhoettes. According to the obtained confsion matrix the accracy of or foregrond estimation algorithm is 98%. The validated corner featres are clstered based on two approaches. When objects move in isolation, an EM algorithm is sed to clster and all the vehicles are tracked. When objects move nder partial occlsion, the corner featres are clstered based on an agglomerative hierarchical clstering approach and the tracking reslt is satisfactory. Since a more reliable estimate of the nmber of objects increases the ability to clster, frther work will be done in this respect in the ftre. If the nmber of objects and the center position estimate of the groped featre points are estimated for each object, the application of EM algorithm cold be sfficient to achieve the desired clstering reqired. As the EM clstering and the agglomerative hierarchical clstering algorithms are implemented off-line, we will contine or work to grop the featre points sing both of these groping approaches nder real time conditions in the ftre. 648

7 SD1.4 REFERE NCES [1] B. Coitinan, D. Beymer, P. McLachlan, and J. Malik, "A real-time compter vision system for vehicle tracking and traffic srveillance," Transportation Research Part C: Emerging Technologies, vol. 6, no. 4, pp , Agst [2] Z. Kim, "Real time object tracking based on dynamic featre groping with backgrond sbtraction." Proceedings of the IEEE Conference on Compter Vision and Pattern Recognition, Jne 28, pp [3] T. Collins, Y. Li, and M. Leordean, "Online selection of discriminative tracking featres," IEEE Transations on Patteren Analysis and Machine Intelligence, vol. 27, no. 1, pp , October 25. [4] N. Bch, J. Orwell, and S. Velastin, "3D extended histogram of oriented gradients (3DHOG) for classification of road sers in rban scenes." Proceedings of the British Machine Vision Conference, September 29, pp [5] J. Shi and J. Malik, "Normalized cts and image segmentation." Proceedings of the IEEE Conference on Compter Vision and Pattern Recognition, Jne 1997, pp [6] A. Yilmaz, O. Javed, and M. Shah, "Object tracking: A srvey," ACM Compting Srveys, vol. 38, no. 4, December 26, article 13. [7] N. Dalal and B. Triggs, "Histograms of oriented gradients for hman detection," vol. 1. Proceedings of the IEEE Conference on Compter Vision and Pattern Recognition, Jne 25, pp [8] S. Birchfield, "Derivation of Kanade-Lcas-Tomasi tracking eqation," Janary [9] J. Shi and C. Tomasi, "Good featres to track." Proceedings of the IEEE Conference on Compter Vision and Pattern Recognition, Jne 1994, pp [1] C. Bishop, Pattern recognition and machine learning. Springer Science Bsiness Media, LLC, 26. [11] R. X and D. C. Wnsch, "Srvey of clstering algorithms," IEEE Transactions on Neral Networks, vol. 16, no. 3, pp , May 25. [12] S. Oh, A. Hoogs, A. Perera, N. Cntoor, C. Chen, c., T. Lee, J., S. Mkherjee, K. Aggarwal, J., H. Lee, L. Davis, E. Swears, X. Wang, Q. Ji, K. Reddy, M. Shah, C. Vondrick, H. Pirsiavash, D. Ramanan, J. Yen, A. Torralba, B. Song, A. Fong, R. Chowdhry, A., and M. Desai, "A large-scale benchmark dataset for event recognition in srveillance video." Proceedings of the IEEE Conference on Compter Vision and Pattern Recognition, Jne 211, pp