MULTI-CAMERA SURVEILLANCE WITH VISUAL TAGGING AND GENERIC CAMERA PLACEMENT. Jian Zhao and Sen-ching S. Cheung

Transcription

1 MULTI-CAMERA SURVEILLANCE WITH VISUAL TAGGING AND GENERIC CAMERA PLACEMENT Jian Zhao and Sen-ching S. Cheung University of Kentucky Center for Visualization and Virtual Environment 1 Quality Street, Suite 800, Lexington, Kentucky ABSTRACT A common goal in many vision alications is to identify and track human objects with distinctive visual features or tags. Examles range from identifying distinct soccer layer by his jersey number to locating the face of an individual that roduces a match in a face recognition system. In this aer, we made two contributions to this visual tagging roblem. First, we roose a general framework for camera lacement. This framework can measure the erformance of any articular camera lacement using simulation method. The otimal lacement strategy can then be obtained by iterative gridbased binary integer rogramming. Second, we focus on tracking secific colored tags used in a rivacy-rotecting visual surveillance network. By building a color classifier for tag detection and using eiolar geometry between multile cameras for occlusion handling, our roosed system can identify, track and visually obfuscate individuals whose rivacy in the surveillance video needs to be rotected. Index Terms multi-camera tracking, camera lacement, eiolar geometry, visual tags, rivacy rotection rotection is articularly challenging as the goal is to obfuscate the images of an individual in ALL CAMERA VIEWS, regardless of whether the small tag is visible to a articular camera. If a tag is visible in only two camera views, its location can be transferred to a third camera view by rojecting the corresonding eiolar lines to the new view as shown in Figure 1. This requires a careful design of distributed algorithms in different cameras so that they can otimally share information about the knowledge of the tags. This is the subject of this aer. Eiolar Lines from two other cameras 1. INTRODUCTION One of the most imortant tasks in distributed camera network is to identify and track common objects across disarate camera views. It is a difficult roblem because image features like corners or scaleinvariant feature transform (SIFT) may vary significantly between different camera views due to disarity, occlusions and variation in illumination. One ossible solution is to utilize semantically rich visual features based either on intrinsic characteristics such as faces or gaits, or artificial marks like numbers on sorts jersey or secial clothing like hats or name tags. We call the roblem of identifying distinctive visual features from an object in two or more camera views the Visual Tagging roblem. Even though visual tagging may require more sohisticated classifiers for tracking or even cooeration from surveillance subjects, it has a wide range of imortant alications. For examle, using distinctive biometric features, visual tagging allows tracking of terrorist susects across a large area like an airort that already has a network of video cameras in lace. Automatic tracking of jersey numbers of layers in a football field can be used to assist coaches in the study of different tactics and strategies. A recent alication of visual tagging is to use secial tags to identify individuals whose rivacy need to be rotected in a video surveillance network [1]. Once a erson is identified to ossess a certain tag, his/her images in all cameras will be obfuscated to rotect the identity. The ideal tag should be small, light and easy to carry. This alication of visual tagging on rivacy Fig. 1. Transfer of tag information via eiolar geometry In this aer, we study various asects of the visual tagging roblem including tagging erformance under different camera lacement, otimal camera lacement strategy, communication strategy for tag localization and its alication in a rivacy rotection system. The main contributions of our aer include the followings: 1. We resent a novel comrehensive visibility metric to measure the erformance of observing a visual tag modeled as a small object with orientation in any arbitrary camera lacement. 2. We develo an iterative otimization method to simultaneously determine the otimal number and ositions of cameras in achieving the desired level of visibility. 3. With colored tags and eiolar geometry between multile cameras for occlusion handling, we are the first to demonstrate how visual tagging between multile cameras can be used in a rivacy rotected video surveillance system. The rest of the aer is organized as follows. In Section 2, we briefly review the state-of-the-art in camera lacement roblem and rivacy rotection schemes. In Section 3, we resent a generic model

2 for measuring the erformance of a articular camera lacement. Section 4 secializes the generic model to define a metric for the visual tagging roblem based on the robability of observing a tag from at least two cameras. Using this metric, we formulate in Section 5 the search of the otimal camera lacement as a Binary Integer Programming roblem. With the otimal camera configuration in lace, we describe how we use visual tagging to enhance rivacy rotection in Section 6. Preliminary exerimental results on both simulations and real videos are resented in Section 7. We conclude the aer by discussing future work in Section 8. others have roosed schemes to obfuscate identity information in video. Datong chen et al.[7] resent a system obscuring the human body while reserving the structure and motion information. Newton et al. develo a face modification algorithm to counter face recognition [8]. Wickramasuri et al. use RFID to track individuals and visually relace them with a static background [9]. Our revious work demonstrate an efficient video in-ainting algorithm to erase individuals for rivacy rotection [10] and resent a data hiding scheme to reserve rivacy information in comressed video [11]. 3. GENERAL VISIBILITY MODEL 2. RELATED WORK In this section, we outline a general model to comute the visibility The roblem of finding the otimal camera lacement has been studied of a single tag P in a confined environment. We assume a two- for a long time. The earliest investigation can be traced back to dimensional model but the analysis can be easily extended to three- the art gallery roblem in comutational geometry. This roblem dimensional. The 2D model, however, is usually adequate for visibility is the theoretical study on how to lace cameras in an arbitrarydimensional. shaed olygon so as to cover the entire area [2]. While the theoretical difficulties of the camera lacement roblem are thoroughly studied, few solutions can be directly alied to realistic comuter vision roblems. Camera lacement has also been studied in the field of hotogrammetry for building the most accurate 3D model. Various metrics such as visual hull [3] and viewoint entroy [4] have been calculations, assuming that the cameras are mounted on the low ceiling tyically seen in many indoor office environment. Given the number of cameras and their lacement in the environment, we can model the visibility of tag P as a ositive function V m(x, y,θ, r), based on the coordinates (x, y) of P, its ose θ with resect to a reference direction, and its half-length r. An examle is shown in Figure 2. develoed and otimization are realized by various tyes of ad-hoc searching and heuristics. These techniques assume very dense lacement of cameras and are not alicable to wide-area wide-baseline camera networks. Recently, Ramakrishnan et al. roose a framework to study the erformance of sensor coverage in wide-area sensor networks [5]. Unlike revious techniques, their aroach takes into account the orientation of the object. They develo a metric to comute the robability of observing an object of random orientation from one sensor, and use that to recursively comute the erformance for multile sensors. While their aroach can be used to study the erformance of a fixed number of cameras, it is not obvious on how to extend their scheme to find the otimal number of cameras as well as how to incororate other constraints such as the visibility from more than one camera. On the other hand, Horster and Lienhart develo a flexible camera lacement model by discretizing the sace into grid and denoting the ossible lacement of camera as a binary variable over each grid oint [6]. The otimal camera configuration is formulated as an integer linear rogramming roblem which can incororate different Fig. 2. General visibility model of a tag P constraints ertinent to a articular alication. While our aroach follows a similar otimization strategy, we adot a more sohisticated robabilistic aroach to cature the uncertainty of object orientation. We also imrove uon the fixed grid oint strategy to rovide more efficient traversal of the search sace. None of the techniques above is suitable for the visual tagging roblem because they do not consider the requirement of viewing an object at arbitrary location and orientation from two or more cameras. The alication of visual tagging in rivacy rotected video surveillance is first roosed by Schiff et al. [1]. They use an Adaboost classifier to identify hard hats and aly articular filtering to track them through time. The rivacy of an individual wearing such a hat is rotected by having his/her face covered by a black box. The choice of hard hats is to rovide a significant target for tracking and recognition and to minimize occlusion. On the other hand, its rominent resence may be singled out in certain environments. In addition, their scheme does not incororate any cues from multile cameras. While Schiff et al. s may be the only scheme that addresses the identification roblem in a rivacy rotected video surveillance, many The visibility function V m rovides an aggregate measure of the rojected sizes of tag P on the image lanes of different cameras. Based on the articular alication, V m can use different aggregation method and incororate a variety of camera and environmental constraints. The secific visibility function suitable for visual tagging will be introduced in Section 4. Note that the deendency of V m on θ allows us to model self-occlusion. It does not, however, model occlusion from other objects and thus we assume that there is only one tag in the environment. While the half-length r of the tag is relatively constant, the coordinates and the ose are random variables governed by a rior distribution f(x, y, θ). This rior distribution can be used to incororate rior knowledge about the environment. For examle, if an alication is interested in locating faces, the likelihood of the head ositions and oses are affected by furnishings and attractions such as television sets and aintings. To correctly identify and track any visual tag, a classification algorithm would require the tag size on the image to be bigger than a certain minimum size, though a larger rojected size usually does

3 not make much difference. Assuming that this minimum size is T ixels, this requirement can be modeled by binarizing the visibility function as follows: { 1 Vm(x, y,θ, r) > T V b (x, y,θ, r T) = (1) 0 otherwise. Finally, we define η, the mean visibility, to be the single metric for measuring the visibility of P over the entire arameter sace: η = V b (x, y, θ, r T) f(x, y,θ)dx dy dθ (2) Excet for the most straightforward environment such as the single camera case discussed in Section 4.1, Equation (2) does not admit a closed-form solution. Nevertheless, it can be easily estimated by using standard Monte-Carlo samling and its many variants. 4. VISIBILITY MODEL FOR VISUAL TAGGING In this section, we resent a visibility model for the visual tagging roblem. This model is a secialization of the general model in Section 3. The goal is to design a visibility function V m that can measure the erformance of a camera lacement for viewing a tag in two or more cameras. In addition to the assumtions listed in Section 3, we further assume that the environment of interest is convex. This imlies that the visibility of the tag with resect to a camera deends only on the coordinates and oses of the tag and the camera. We also assume the basic inhole camera model. Let us start with the simle case for one camera Visibility for single camera Recall the visibility function V m(x,y, θ) measures the rojected size of the tag on the image lane. Instead of arbitrarily choosing a coordinate system and a reference direction, we directly comute the rojected size l based on the following geometrical quantities of the tag P and the camera C as well as the otical roerties of C: d α β f FOV the length of line segment PC that joins the camera inhole to the center of the tag the angle between the tag orientation and PC the angle between the camera rojection direction and PC the camera s focal length the camera s field of view the width of a ixel on the image lane These quantities are illustrated in Figure 3. For convenience, we assume that the image lane is distance f in front of the camera inhole. First, it is straightforward to see that P is visible by the camera if and only if the following two conditions hold: 1. P is within the camera s field of view or β < FOV/2 (3) 2. P is not self-occluded when seen from the camera or α < π/2 (4) Fig. 3. The size of object of interest in the image, l, is determined by the object size r, distance d, and angles α, β We thus define the visibility V (P, C) between the tag P and a single camera C as follows: { l α, β satisfy conditions (3) and (4) V (P, C) = 0 otherwise ( π ) ( ) FOV = U 2 α U β l (5) 2 where U(.) is the unit ste function. A threshold version is sometimes more convenient: { 1 if V (P, C) > T V b (P, C T) = (6) 0 otherwise Second, we exress the rojected tag size l in terms of the basic geometric roerties defined above. Define 1 and 2 as the angles between PC and the lines joining either ends of the tag to the inhole. We have four different cases: Case 1: β and 2 are on the same side with β > 2. This is illustrated in Figure 4(a). Using the two right-angled triangles formed among the inhole, the otical center and the two resective endoints of the rojection, we can comute l in terms of the number of ixels as follows: l = [tan(β + 1) tan(β 2)] f Case 2: β and 1 are on the same side with β > 1. This is illustrated in Figure 4(b). Based on a similar argument as case 1, we have l = [tan(β + 2) tan(β 1)] f Case 3: β and 1 are on the same side with β < 1. This is illustrated in Figure 4(c). Based on a similar argument as case 1, we have l = [tan(β + 2) + tan( 1 β)] f (7) (8) (9)

4 (a) case 1 (b) case 2 (c) case 3 (d) case 4 Fig. 4. Four cases to comute the rojected tag size l. Case 4: β and 2 are on the same side with β < 2. This is illustrated in Figure 4(d). Based on a similar argument as case 1, we have l = [tan(β + 1) + tan( 2 β)] f (10) Note that Equation (7) is equivalent to Equation (10), and Equation (8) is equivalent to Equation (9). The two sets of equations can be differentiated based on whether α and β are on the same side of PC. In addition, we can exress the tangent of 1 and 2 in terms of the tag arameters α, β, r and d as follows: tan 1 = r cos α d + r sin α (11) and tan 2 = r cos α (12) d r sin α Combining the above observations with Equation (11) and (12), we can comute l as a function of α, β, r and d: l(α, β, d, r) = 2dr cos α(tan 2 β+1) (f/) d 2 r 2 sin 2 α+r 2 tan β(tan β cos 2 α sin2α) 2dr cos α(tan 2 β+1) (f/) d 2 r 2 sin 2 α+r 2 tan β(tan β cos 2 α+sin2α) α, β on the same side (13) of PC otherwise 4.2. Visibility for Visual Tagging To extend the single-camera case in Section 4.1 to multile cameras, we note that the visibility of the tag from one camera does not affect the other and so each camera can be treated indeendently. The visual tagging roblem requires that the tag must be visible by at least two cameras. Given N cameras C 1, C 2,..., C N, we define the visibility function V m(x, y,θ, r) for visual tagging to be the second largest rojected tag size among all the cameras: V m(x,y, θ, r) = max V (P, C i) (14) i {1,2,...,N},i k where V (P, C i) is the visibility of tag P with resect to camera C i as defined in Equation (5) and C k is the camera that catures the largest tag image or k = arg max j {1,2,...,N} V (P, C j). Even if the environment is densely covered with cameras, there is no guarantee that a tag at an arbitrary osition will be visible to at least two cameras a tag next to and facing the wall is only visible if there are two cameras right in front of the tag. In the actual design of camera networks, we would like to avoid such athological cases and to adot the design if most of the environment is erfectly visible. We call the area in the environment a Perfect Zone in which a tag of half-length r, regardless of its ose, is visible to two cameras. In

5 other words, the Perfect Zone can be defined as Perfect Zone = {(x, y) : V m(x, y, θ, r) > 0 for all θ} = {(x, y) : V b (x, y,θ, r T) = 1 for all θ} (15) 5. OPTIMAL CAMERA PLACEMENT In the revious sections, we show how to comute visibility of arbitrary camera lacement. In this section, we demonstrate how to comute the otimal camera lacement the minimum number of cameras used, their oses, and their ositions in the environment in order to achieve a target η t. Due to the difficulty in obtaining a continuous solution over an arbitrary-shae environment, we follow a similar aroach as in [6] by finding an aroximate solution over a discrete grid. We first discretize the environment into a lattice gridp of N grid oints {P i : i = 1, 2,..., N } where we are interested in finding the tag visibility. We also discretize the camera sace, that includes both the 2D locations and the orientation, into a lattice gridc of N c grid oints {C i : i = 1, 2,..., N c}. To formulate the otimization roblem, we associate each camera grid oint C i with a binary variable b i such that { 1 a camera is resent at Ci b i = (16) 0 otherwise The otimization roblem can be described as the minimization of the number of cameras: N c min b i (17) i=1 subjected to the visual tagging constraint: N c b i V b (P j, C i T) 2 (18) i=1 for each ossible tag configuration P j and the single camera constraint: b i 1 for each camera location (x, y) (19) all C i at (x,y) The first constraint (18) reresents the requirement of visual tagging that all tags must be visible by at least two cameras. As defined in Equation (6), V b (P j, C i T) measures the visibility of tag P j with resect to camera at C i. In other words, P j satisfying the constraint (18) must be in the erfect zone. The second constraint in (19) is a set of inequalities to guanratee that only one camera is laced at any 2D location. The otimization roblem in (17) with constraints (18) and (19) forms a standard Binary Integer Programming (BIP) roblem. While the general BIP roblem is NP-hard, fast aroximate solutions exist and can be obtained using software libraries such as l solve [12]. The choice of the grid oints in gridp and gridc affect the outcome of the otimization. As discussed in Section 4.2, there is no guarantee that a tag at a random location can be visible by two cameras even if there is a camera at every camera grid oint. Thus, the otimization roblem may not have a solution if the tag grid oints are randomly laced. Instead, we start our search rocess at a configuration that guarantees a solution and then traverse the search sace by gradually increasing the density of both gridp and gridc. The initial configuration of gridp is to lace a single tag at the center of the largest circle inscribed within the environment. This tag is guaranteed to be in the erfect zone if we ut five center-facing cameras equally saced on the circle no matter what the tag orientation is, the tag will always be visible to two or more cameras. In order to roduce a better estimate of η, gridp then grows uniformly in density within the interior of the environment but remains at least one interval away from the boundary. gridc maintains its initial density until the BIP solver fails to return an answer, at which oint the density of gridc is increased. The iteration terminates when the target η t is achieved or the density of gridc exceeds a redefined limit. The above rocess is described in Algorithm 1. The algorithm is guaranteed to terminate by setting a very high maxdensity so that the entire environment can be covered with cameras. In ractice, as we will show in Section 7, it only requires a few iteration to arrive at a very high level of η. Inut: initial grid oints for cameras gridc and tag gridp, η t, maximum grid density maxdensity Outut: Camera lacement camp lace Set η = 0; while η η t AND density(gridc) maxdensity do foreach C i in gridc do foreach P j in gridp do Calculate V b (P j, C i 0); end end Solve newcamplace = l solve(gridc,gridp, V b ); if newcamplace==empty then Increase density of gridc; Decrease density of grip ; else camp lace = newcamp lace; end Calculate η for camplace by Monte Carlo Samling; Increase density of gridp ; end Algorithm 1: Otimal camera lacement algorithm 6. VISUAL TAGGING FOR PRIVACY PROTECTION In this section, we describe a system that uses visual tagging to rotect rivacy of selected individuals in a multi-camera video surveillance network. The cameras are ositioned based on the otimal lacement strategy described in Section 5. All cameras are calibrated such that camera C k knows the set of fundamental matrices F ik from camera C i to C k. In other words, if x i (in homogeneous coordinate) is a oint on the image lane of C i, then l k = F ik x i is the eiolar line to x i on the image lane of C k. Individuals whose rivacy need to be rotected are wearing small rectangular colored tags. Each tag has an unique color for which we have reared a color classifier. Our current imlementation uses a re-trained Gaussian Mixture Model classifier on the hue and saturation of each color. Using these classifiers, each camera identifies all ixels that match these colors, erforms comonent grouing on ixels with the same color, and comutes the centroid of each grou. The image coordinates of the centroids and their corresonding colors along with the camera s own ID are broadcast to all other cameras.

6 After receiving all messages from its eers, each camera C k decodes the messages and builds a contingency table between cameras and color tags. If more than one camera rovide information about a tag with a secific color, camera C k comutes the corresonding eiolar lines and estimates their oint of intersection on its own image lane. Since these eiolar lines must come from the same tag, they will intersect at a oint. By cross-correlating eiolar lines from other cameras, it is ossible for camera C k to identify the image location of these virtual tags even though they are not directly observed by C k. In ractice, these eiolar lines intersect in a small region instead of a single oint due to the uncertainty in identifying the rojection of the centroid from each camera and in comuting the fundamental matrices between wide baseline cameras. While this error can be reduced by taking advantage of the secific shae of the tags, it is usually tolerable as it will be further correlated with the identified moving objects as discussed in the sequel. Based on our earlier work in [10], each camera combines a background subtraction algorithm with a robabilistic tracer to motionsegment individual objects in the video. Objects overlaed with any directly-observed tags or virtual tags are erased from the video by using an efficient object-temlate based in-ainting scheme [10]. Preliminary results of this system are shown in Section EXPERIMENTAL RESULTS In this section, we reort three sets of exeriments. In the first set of exeriments, we aly the algorithm introduced in Section 5 to derive the otimal camera lacement for a target mean visibility η t. To facilitate testing of the camera lacement strategy, our second set of exerimental results are based on simulating a virtual environment in which cameras can be laced in arbitrary ositions. Finally, we illustrate the use of visual tagging for rivacy rotection in an actual three-camera surveillance network. All the simulations assume a room of dimension 10m 10m and a tag of half-length r = 10cm long. For the camera and lens models, we assume a ixel width of 5.6 µm, focal length of 8 cm and the field of view of 120 degrees. The threshold T for visibility is set to 5 ixels. By setting the target mean visibility η t = 0.95, the otimal lacement algorithm terminates after five iterations. The snashots after the first, third and fifth iteration are shown in Figure 5(a) to 5(f). Figure 5(a) to 5(c) show the tag grid oints (hollowed circles), the camera grid oints (solid circles) and the resulting otimal camera ositions and oses (solid black circles with camera orientations). Both the camera and tag grids are refined over the course of these iterations. Figure 5(d) to 5(f) show the corresonding visibility function V m(x, y,r, θ) at each of the random samle location (x, y), averaged over all ossible θ under an uniform distribution. A brighter ixel indicates a higher average visibility and a white ixel belongs to the erfect zone that it is visible to two or more cameras regardless of the orientation. The mean visibility η is comuted based on these random samles. Notice that while both the room and the grids are symmetric under 90 o rotations, the otimal solution at each stage does not ossess such symmetry. The reason is that our software randomly icks one of the many otimal configurations that satisfy the otimality criteria. To validate the obtained otimal camera lacement, we simulate a 3-D environment of the same dimension in OenGL and ut twelve cameras according the otimal lacement comuted in Figure 5(c). A humanoid wearing a visual tag with random osition and ose is created [13]. A samle of camera views are shown in Figure 6. Out of 140 random humanoid scenes, our visual insection identifies 14 scenes in which the tag is not visible to at least two cameras. This re- (a) Iteration one (b) Iteration three (c) Iteration five (d) η = (e) η = (f) η = Fig. 5. First few iterations of the camera lacement algorithm. Figures 5(a) to 5(c) show the grid oints as well as the otimal camera oses and locations. The tag grid consists of only the hollowed circles and the camera grid consists of all the solid grid oints. The black dots show the otimum camera osition after the iteration and the arrows show the camera ose. The corresonding average visibility erformances are shown in Figure 5(d) to 5(f). sults in an estimate of η 0.9, lower than the exected Part of the reasons is that our model does not take into account the elevation of the camera, which makes the tag invisible when the humanoid is too closed to the camera. However, if we restrict the ositions of the humanoid to be within the middle 7.5m 7.5m area of the room, we obtain only 1 miss out 150 random scenes. This confirms the resence of the erfect zone in the middle of the environment. The final exeriment involves using three real cameras mounted on the ceiling of our laboratory. Figure 7(a) to 7(c) show the three views of the same erson. The foreground moving blobs are shown in Figure 7(d)) to 7(f). Note that the tag is only visible in Cam1 and Cam2 but not in Cam3. Figure 7(g) and 7(h) show that our classifier can clearly identify the tag in Cam1 and Cam2. No such ixels are detected in Cam3, which resorts to calculating the two eiolar lines in Figure 7(i) based on information sent by Cam1 and Cam2. For Cam1 and Cam2, the tag ixels overla with those of the foreground blobs. As a result, the foreground blobs are erased and the outut frames are shown in Figure 7(j) and 7(k). As for Cam3, since the intersection of the two eiolar lines lie within the foreground blob, the blob is also erased as shown in Figure 7(l). 8. FUTURE WORK In this aer, we have roosed a multi-camera surveillance system with visual tagging and camera lacement model. By building a camera lacement metric using lanar geometry, we have derived an otimal camera lacement strategy using iterative grid based binary integer rogramming. Equied with the otimal camera lacement, we have resented a multi-camera surveillance system caable of robustly identifying the visual tag and rotecting the rivacy of selected individuals by obfuscating their images in all camera views. We are currently extending the camera model to handle mutual occlusion among multile tags so that robust tracking in crowded en-

7 (a) camera 1 (b) camera 2 (c) camera 3 (a) Cam1 original (b) Cam2 original (c) Cam3 original (d) camera 4 (e) camera 5 (f) camera 6 (d) Cam1 foreground (e) Cam2 foreground (f) Cam3 foreground (g) camera 7 (h) camera 8 (i) camera 9 (g) Cam1 tag (h) Cam2 tag (i) Cam3 eiolar lines (j) camera 10 (k) camera 11 (l) camera 12 (j) Cam1 outut (k) Cam2 outut (l) Cam3 outut Fig. 6. Different views of the virtual environment. The cameras are ositioned as in Figure 5(f) with camera 1 being the one at the lower left corner and the camera number increases counterclockwise around the erimeter of the room. Fig. 7. Oututs of the rivacy rotection system. 7(a) to 7(c) are the originals and 7(d) to 7(f) are the foreground blobs. 7(g) and 7(h) are the detected tags and 7(i) are the two eiolar lines comuted based on the ositions of the tag from other cameras. The outut in-ainted frames are shown in 7(j) to 7(l).

8 vironment can be realized. We are also imroving the registration techniques between different camera views by combining all available cues into a robabilistic framework. 9. REFERENCES [1] J. Schiff, M. Meingast, D. Mulligan, S. Sastry, and K. Goldberg, Resectful cameras: Detecting visual markers in realtime to address rivacy concerns, in International Conference on Intelligent Robots and Systems (IROS), [2] Joseh O Rourke, Art Gallery Theorems and Algorithms, Oxford University Press, [3] D. Yang, J. Shin, A. Ercan, and L. Guibas, Sensor tasking for occuancy reasoning in a camera network, in 1st Worksho on Broadband Advanced Sensor Networks (BASENETS). IEEE/ICST, [4] P.-P. Vazquez, M. Feixas, M. Sbert, and W. Heidrich, Viewoint selection using viewoint entroy, in Proceedings of the Vision Modeling and Visualization Conference (VMV01), [5] S. Ram, K. R. Ramakrishnan, P. K. Atrey, V. K. Singh, and M. S. Kankanhalli, A design methodology for selection and lacement of sensors in multimedia surveillance systems, in VSSN 06: Proceedings of the 4th ACM international worksho on Video surveillance and sensor networks, New York, NY, USA, 2006, , ACM Press. [6] E. Horster and R. Lienhart, On the otimal lacement of multile visual sensors, in VSSN 06: Proceedings of the 4th ACM international worksho on Video surveillance and sensor networks, New York, NY, USA, 2006, , ACM Press. [7] D. Chen, Y. Chang, R. Yan, and J. Yang, Tools for rotecting the rivacy of secific individuals in video, EURASIP Journal on Advances in Signal Processing, vol. 2007,. Article ID 75427, 9 ages, 2007, doi: /2007/ [8] E. N. Newton, L. Sweeney, and B. Main, Preserving rivacy by de-identifying face images, IEEE transactions on Knowledge and Data Engineering, vol. 17, no. 2, , February [9] J.Wickramasuri et al., Privacy rotecting data collectino in media saces, ACM Multimedia, , October [10] S.-C. Cheung, J. Zhao, and V. Venkatesh M., Efficient objectbased video inainting, in Proceedings of IEEE International Conference on Image Processing, ICIP 2006, 2006, [11] W. Zhang, S.-C. Cheung, and M. Chen, Hiding rivacy information in video surveillance system, in Image Processing, IEEE International Conference on. IEEE, [12] Introduction to l solve , htt://lsolve.sourceforge.net/5.5/. [13] K. Agarway and P. Winston, Walker, htt://