Your SlideShare is downloading. ×
  • Like
50120140502009
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×

Now you can save presentations on your phone or tablet

Available for both IPhone and Android

Text the download link to your phone

Standard text messaging rates apply
Published

 

Published in Technology
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Be the first to comment
    Be the first to like this
No Downloads

Views

Total Views
74
On SlideShare
0
From Embeds
0
Number of Embeds
0

Actions

Shares
Downloads
2
Comments
0
Likes
0

Embeds 0

No embeds

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
    No notes for slide

Transcript

  • 1. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING & ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME TECHNOLOGY (IJCET) ISSN 0976 – 6367(Print) ISSN 0976 – 6375(Online) Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME: www.iaeme.com/ijcet.asp Journal Impact Factor (2014): 4.4012 (Calculated by GISI) www.jifactor.com IJCET ©IAEME CROWD BEHAVIOUR ANALYSIS CONSIDERING INTER-PERSONNEL ACTIVITIES IN SURVEILLANCE SYSTEMS Najmuzzama Zerdi1, Dr. Subhash S Kulkarni2, Dr. V. D. Mytri3, Kashyap D Dhruve4 1 2 (Professor and Head, Dept. of E&C, K.C.T.E.C Gulbarga- 585101, Karnataka, India) (Professor and Head, Dept. of E&C, PES Institute of Technology - Bangalore South Campus, Bangalore India) 3 (Principal, AIET, Gulbarga, Karnataka, India) 4 (Technical Director, Planet-I Technologies, Bangalore, India) ABSTRACT In public and private places, people panic when unpredicted events occur. The modelling and analysis of the crowd surveillance videos is an issue that exists due to the unconstrained behaviors observed. The use of computer aided techniques to address this issue has been proposed by researchers. Similarly, we proposed a Graph theoretic approach based Crowd Behavior Analysis and Classification System (GCBACS) model in this paper. The crowd personnel behavior is observed and graph theoretic approach is used for analysis and classification of the activity observed. Streak Flows is considered to observe crowd behaviors. The novelty of GCBACS is that considers inter personnel activities for analysis and classification. The GCBACS is evaluated using varied surveillance videos in the experimental study. The results presented prove that the GCBACS is able to achieve accurate classification results in dense and sparse crowd scenarios. The GCBACS is compared with the Spatio-Temporal Viscous Fluid Field method. The results prove that the GCBACS outperforms the Spatio-Temporal Viscous Fluid Field method. Keywords: Video Surveillance, Crowd Behavior, Crowd Motion, Optical Flow, Streak Lines, Streak Line Flow, Path Lines, Threshold, Graph Theory. 1. INTRODUCTION Now days video surveillance systems are fed by multiple high definition video streams, they have become a very common feature in public and private spaces. The video surveillance systems can be implemented manually and the large volume of data captured from high-throughput 71
  • 2. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME surveillance system make manual analysis extremely time consuming, prone to human error. The steady population growth observed in the past decade have resulted large crowd movements especially in public spaces like airports , train stations , bus stations, shopping malls, religious places, etc. The number of crowd accidents observed have increased in the recent times [1]. The automated system, which is used to classify the movements of crowds or detect abnormal activity, can be considered as an open research issue. A crowd can be considered as a collection of people distributed over the region of interest and tracking of human activity or personnel counting within video surveillance systems has been researched upon [3-4] for some time now. The open research issues that exist and require attention with respect to crowd analysis can be listed as modeling or knowledge extraction from crowd patterns [5-8] and crowd behavior analysis [9-11]. Limited work is carried out to classify the behavior of crowds in surveillance systems. The research work presented in this paper introduces the Graph theoretic approach based Crowd Behavior Analysis and Classification System ( ). To achieve accurate classification results the behavior of the personnel in the crowd needs to be analyzed first. The behavior of the personnel in the crowd can be analyzed based on the motion or trajectory activities observed. Based on the behavior of the personnel analyzed, it can be classified into normal or abnormal activity. Abnormal activity detection is achieved by observing unusual behavior of personnel or group of personnel within a crowd. Abnormality detection refers to the detection of unusual behavior of individuals or a group in a crowd scene and the Activities like instantaneous disbursement, sudden convergence or fighting are classified as abnormal activities. To analyze the behavior of personnel in the crowd, tracking methodologies are generally used. The commonly used tracking methodologies [2] [3] [4] implemented and it is fail when large crowds are considered. To overcome this drawback, researchers proposed the consideration of fixed cell sizes to identify local trajectories and later map it together to detect the personnel trajectory patterns [8] [10] [12]. However, the frames are split into uniform cells in these approaches. The use of optical flow techniques within each cell is considered to obtain the trajectory patterns of personnel within a cell. For instance, uses the optical flow techniques exhibited better results when compared to traditional tracking methodologies [13]. The drawback of the optical flow is that only two consecutive frames are considered to obtain personnel trajectory patterns. The optical flow method are not able to capture long term does temporal dependencies [14]. To overcome this drawback the concept of particle flow was introduced in [7]. The particle flow computation is achieved by displacing a grid of particles with optical flow through numerical integration techniques, which providing trajectories that relate a particles original position to its position at a later time. The particle flow mechanisms proved to be computationally very heavy and minute personnel motion details were ignored. The introduction of streak lines flows to obtain the trajectories of personnel in the crowd proved to provide accurate analysis results [13]. For crowd behavior classification in [15] an unsupervised machine learning technique based framework was proposed. The framework in [16] considered hierarchical Bayesian models to connect the visual features, “atomic” activities and the interactions for classification. In [12] streak-lines coupled with social force models were used to detect abnormal activities. The work carried out so far by researchers, primarily concentrates on analysis of activities amongst a few personnel present in the crowd only, and do not take into account the inter personnel activities for classification. To overcome this drawback the presented in this paper which considers inter personnel activities for analysis. The inter personnel activity consideration, enables analysis in dense and sparse crowd scenarios. The inter personnel activities are monitored through the motion vectors observed. To acquire the behavioral vectors of personnel in the crowd video an optical flow is initially computed. Based on the optical flow the path lines and streak lines are presented here. However, the path lines, streak lines are used to derive the streak flow vectors which 72
  • 3. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME define the potential and personnel flow in crowds. Every frame of the video is analyzed using graph theoretic approaches and it is also considers simultaneously density, direction, velocity, focuses analysis on specific regions where the density of the motions is high. The GCBACS considers each frame as a graph with sub graphs. All the frames are analyzed and the cumulative variance is computed. If the GCBACS observes that the cumulative variance is greater than a threshold the activity of the personnel in the crowd is classified as an abnormal activity. The remaining manuscript is organized as follows. Section two discusses the related work. The GCBACS is discussed in section three of the paper. The experimental study conducted to evaluate the performance of the GCBACS is discussed in section four of the paper. The conclusion is discussed in last section in this paper. 2. LITERATURE SURVEY In this section of the paper a brief of the literature review conducted during the course of the research work presented here is discussed. Most related works focus on the analysis of activities introduced by some individuals between them. Myo Thida et al. [6] proposed a manifold learning-based framework for the detection of anomalies that occur in a crowded scene. The spatiotemporal LE method is employed to study the local motion structure of the scene. The pair-wise graph is constructed by considering the visual context of multiple local patches in both spatial and temporal domains. The drawbacks of this paper related to localization of the abnormal regions. Kristjan Greenewald et al [7] proposed to learn the normative multi-frame pixel joint distribution and detect deviations from it using a likelihood based approach. Due to the extreme lack of available training samples relative to the dimension of the distribution, a mean and covariance approach and consider methods of learning the spatio-temporal covariance in the low-sample regime. The space-time pixel covariance for crowd videos can be effectively represented as a sum of Kronecker products using only a few factors, when adjustment is made for steady flow if present. Yong Wang et al. [8] represents an abnormal behavior detection approach which is convenient and available for camera sensor networks. Trajectory analysis and anomaly modeling are carried out by single-node processing; therefore, the anomaly detection is performed by multimode voting. A first-order Markov model is used to build the anomaly transition matrix and the detection threshold can be determined by training the obtained symbol sequences. The voting mechanism reaches a final decision in accordance with local decisions of camera nodes. The drawbacks of this method that it is undesirable for complex abnormal behavior detection to only consider the trajectory information. Xiaobin Zhu et al.[9] proposed a weighted interaction force estimation in the social force model(SFM)-based framework, in which the properties of surrounding individuals in terms of motion consistence, distance apart, and angle-of-view along moving directions are fully utilized in order to more precisely discriminate normal or abnormal behaviors of crowd. To avoid the challenges in object tracking in crowded videos, here they perform particle advection to capture the continuity of crowd and use these moving particles as individuals for the interaction force estimation. the properties of surrounding individuals in terms of motion consistence, distance apart, and angleof-view along moving directions are fully explored, in order to more precisely discriminate normal or abnormal behaviors of crowd. 73
  • 4. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME 3. GRAPH THEORETIC APPROACH BASED CROWD BEHAVIOR ANALYSIS AND CLASSIFICATION SYSTEM (GCBACS) 3.1 SYSTEM PRELIMINARIES model overview Figure 1: We present a surveillance video .Let us consider, the video represents a set of frames and the dimension of each image is pixels. A frame at the time instance and . Similarly the frame at the time instance is represented as .The frame is split into a number of blocks and a mesh based structure is created for computational ease. Let the set represent the crowd personnel to be observed in the surveillance video space . The set consists of personnel. The trajectory of the personnel i.e. at the time instance can be represented as . At the initial instance i.e. the trajectory is represented as . Generally, the trajectories is used for optical flow computations. From the optical flows we can observed that the streak lines and path lines of the personnel in the crowd are is derived which is used for analysis. For computed frame wise. Therefore, the streak flow analysis a graph theoretic approach is adopted in the graph theoretic approach based Crowd Behavior Analysis and Classification System GCBACS. We denote each frame as a graph and analysis is carried out based on the similarity and deviations are noticed. In order to effectively recognize the Cumulative variance is computed considering all the previous frames and the current frame. Therefore, if the variance is greater than the threshold then abnormal activity is said to be modeling is illustrated in Figure 1. detected. The proposed 3.2 OPTICAL FLOW ESTABLISHMENT In GCBACS the Lucas & Kanade based methodology is used to compute the differential optical flow of the crowd vectors. The optical flow enables trajectory detection of personnel in the crowd. Let the velocity field defined over the set be represented as . The velocity satisfy the continuity in time and continuity in space domain to obtain smooth optical flows. To achieve optical flow computation a hierarchical graph structure is considered to represent the video . 74
  • 5. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME Let the levels of the graph be defined as . If represents the velocity then the optical flow residual vector is used to minimizes the function vector . Similarly the matching function can be minimized using the residual vector . The primary guess for the level of the optical flow is denoted as . The value of is obtained by optical flow computations from to . The frame and can be represented on the basis of the optical flows computed at all the levels and is represented as From the above equation, we can denoted as, are two integer values and and represent the previous two frames. Based on the above equations it can be acquired that there exist a and . The optical flow representation of the domain definition difference between which is defined over the window size instead of frame using . Let us Consider that the displacement vector is and image position vector is .the matching function is minimizes by the vector .The function is represents as An iterative Lucas –Kanade method is adopted to solve the function From the first order expansion about the point can described as Therefore, the point is described as and is represented as instead of is the temporal frame image derivative at the point 75 Equation 4 . The
  • 6. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME The derivative at is denoted as From Equation 5we can denote , from The derivatives in the the next image Based on and is the gradient vector and can be defined as and can be evaluate directly from the image , which is a neighborhood of the point independently . The derivative images satisfy the expression and can be represents as defined above the computation of Equation 12 can be further simplified and written as: Where 76 is given as
  • 7. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME As and matrix is invertible, then the optical flow vector Based on the above equation we can observed that is defined as , it is evident that contains information of the gradients in the and direction of . A large number of iterations have to be represent considered to obtain accurate optical flow of personnel in the crowd video frames. Let the number of iterations required and . Based on the optical flow computations from the initial guess for pixel displacement is obtained. The initial guess is given as . If represents the new image based on , provided then Therefore, using optical flow methodology in and mismatch vectors are obtained. The residual vector function , defined as Using Equation 14 the solution of Where the mismatch vector matrix In Equation 16 the residual pixel trajectory vector minimizes the error represents the can be obtained and is defined as is given as frame difference and is defined as 77
  • 8. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME From Equation 18, we can observe that the spatial derivatives and are computed only once initially and is constant for an entire iteration loop. The parameter vector is iteratively steps. That vector is the amount of residual difference between the video frames computed at after translation by the vector . Based on the matrix and , is computed. The new pixel displacement is that is computed in the step and is defined as is below a preset The iteration steps to achieve convergence are repeated till the value of threshold or the number of maximum iterations are completed. If represents the number of iterations required to reach convergence then the optical flow vector can be defined as The optical flow based on the velocities in the and be represented as direction at the time instance can 3.3 STREAK-LINE FLOW ESTABLISHMENT First, we introduced a streak-line representation of flow based on the optical flow and optical flow computed present gaps in the trajectories of personnel in the crowd. In GCBACS the gaps in the optical flow of similar motion vectors are filled using the streak lines [15] and new trajectory vectors are established prior to analysis using graph theoretic approaches. Let us consider a particle at position in the tth time instance, present in the frame and it is represented as . The advection of the particle is achieved by are obtained from the optical flow vectors. For all the frames and time using particle advection we can acquire a vector matrix. The columns of the matrix can be used to obtain the particle trajectory details from time to the current time that is described as path lines. The is introduced to represent the path lines. The row of the matrix can be used to define the streak lines that connect the particles from video frames that generated from the position . In this paper, the streak lines is represented using notation Inconstancies are identified in the streak lines. The drawback in [15] the extended particle was introduced based on the position and the optical flow velocities can be overcome by proposed method. The th extended particle can be denoted as 78
  • 9. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME Where, and . As per the streak lines the behavior of the personnel is obtained. Using the streak lines, the streak flow is computed and is defined as Therefore, the streak line computation is realized by integrating the optical flows computed have to be computed. The computation of and forming extended particles. To compute , and and are similar in nature. If then let us consider a vector to obtain the has three pixels as its neighbors which forms streak flow in the direction. The extended particle a triangle. is considered as the interpolations of the neighboring pixels and is represented as From the Equation 26 identify the index of the pixel, and Using the interpolation method the parameters vectors in and based on Equation 26 we can state Then, are the elements of the matrix represents the basis function. are obtained. For all the . Therefore, following equation, 25 and 26 similarly can be computed. Using and the streak flow is acquired. In Graph theoretic model use of streak approach based Crowd Behavior Analysis and Classification System flow to acquire the trajectory of the personnel in the crowd is considered as the streak flow methodology enables instantaneous change observation when compared to particle flows. 3.4 GRAPH THEORETIC APPROACH FOR ANALYSIS AND CLASSIFICATION In GCBACS model the behavior of crowd personnel noticed using the streak flow is analyzed using a graph structure, which is adopted for all the frames of the video. Let us consider is a constructed graph obtained from the streak flow . The vertices of the graph are the number of pixel vectors noticed and represents as Therefore, the edges of the graph can be represented as Where, streak flow computed represents a planar field, and based on the decomposition obtained by Helmholtz. is the irrotational part of the vector field and is the 79
  • 10. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME incompressible part. In [22] two functions are introduced such that, .However, the stream function and the velocity potential function are described using Fourier Transforms introduced in [22]. The functions and are represent as Then it follows immediately that the function contribute details of the steady motion vectors and contributes details of the random motion changes detected. Therefore, By combining the and vectors the potential functions of the video frame is computed and the edge set can be represented as Therefore, an abnormal event can be detected from the consecutive frames is considered i.e. graph and graph the relation between the graphs can also be examined as the relation amongst the sub sets of and and is represented as represents the number of vectors Based on the above equation we can denoted that common to the graphs and . The number of nonaligned graph nodes in the two frames is represented by Therefore, is the number of matched sub graphs identify in the graph frames and . represents the sub sets of . represents the sub sets of The matching of the energy potential functions of the frames is represents as 80
  • 11. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME However, the local differences in the sub sets are measured to observed the finer movements in the graphs and is computed using is a function that defines the matching between the sub During the process sets and . Combining all the above definitions the variance between two frames represented as graphs and calculated as Where be given as is a predefined integer and the cumulative variance noticed till the frame can Therefore, if the value of the cumulative variance is greater than a predefined threshold then abnormal event is identified in the video and it is assigned the class 1 else 0. The classification can be described as Having discussed the modelling of crowd behavior using streak flows and analysis using graph theoretic approaches in the , the next section of this paper presents an experimental study considering the proposed system. 4. EXPERIMENTAL STUDY AND PERFORMANCE EVALUATION The effectiveness of the performance of GCBACS model is evaluated on the data set from University of Minnesota [23] and a web data set. The video sequences [23] consists of abnormal and normal crowd personnel videos and is referred to as Case 1. The web data set consists of a video sequence of two persons and is referred to as Case 2 hereafter. In Case 2 a person pushes the other person from behind and then runs. The purpose of the GCBACS model is to detect the abnormal behavior in Case 1, i.e. sudden dispersal of crowd personnel and the falling of the person in Case 2. The performance of the GCBACS is compared with the Spatio-Temporal Viscous Fluid Field ( ) 81
  • 12. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME method proposed in [17]. The Matlab platform is used to develop GCBACS model. To represent the recognition results the authors of this paper have considered the use or bar charts. The green section represent normal crowd activity and the red section represent abnormal crowd activity. The use of receiver operating characteristic curves (ROC) [24] has been considered to evaluate the classification performance. The area under the ROC curve and classification error plots are considered to evaluate the performance of GCBACS and Spatio-Temporal Viscous Fluid Field (VFF) method. 4.1 CASE 1 :STUDY USING UNIVERSITY OF MINNESOTA DATASET Figure 2: Comparison results of crowd behavior recognition and classification for videos in Case 1 Using and . The ground Truth is also shown The recognition results obtained, considering GCBACS and VFF for Case 1 is shown in Figure 3. From the figure it is clear that the GCBACS achieves better activity classification when is 0.18 and for GCBACS is 0.5. compared to the VFF method. The misclassification ratio for The unconstrained movement of people in the crowd observed by the Spatio-Temporal Viscous Fluid Field (VFF) method resulted in a larger number of misclassified frames seen by the small red intermediate regions. The misclassification ratio of is lower as inter personnel activities are considered for classification. The ROC curve obtained is shown Figure 3. The curves prove that the GCBACS accurately classifies and analyzes the crowd behavior in Case 1 when compared to the scheme. The area under the ROC curve considering GCBACS was found to be 0.88 and for 0.67 and is shown in Figure 4. The average classification error considering GCBACS was found to be 0.17. In the case of the Spatio-Temporal Viscous Fluid Field (VFF) the average classification error is 0.29. The results obtained for the classification errors observed is shown in Figure 5. 82
  • 13. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME Figure 3: Curves for Crowd Activity Classification based on Figure 4: Area under Curve considering Figure 5: Classification Error considering 83 and and and for Case 1 for Case 1 for Case 1
  • 14. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME 4.2 CASE 2 : STUDY USING WEB DATASET In Case 2 the surveillance video consists of a sparse crowd scenario i.e. only two people. In such conditions the interpersonal behavior capture is critical to achieve accurate analysis. The recognition results obtained considering GCBACS and VFF are shown in Figure 6. From the figure it is clear that the proposed GCBACS exhibits higher efficiency to capture interpersonal behavior when compared to the Spatio-Temporal Viscous Fluid Field (VFF) method. This is also proved by the fact that a lower number of false positives are seen in the bars. The misclassification ratio is 0.22 and 0.03 for GCBACS and VFF. The classification accuracy of the GCBACS when compared to the VFF is proved by the ROC curve plot shown in Figure 7. The area under the ROC curve is around 0.98 for GCBACS and 0.73 for VFF. The area under the ROC curve graph is shown in Figure 8. The classification error observed is shown in Figure 9 of this paper. The average classification error for is 0.24 and for is 0.09. Figure 6: Comparison results of crowd behavior recognition and classification for videos in Case 2 Using and . The ground Truth is also shown Figure 7: Curves for Crowd Activity Classification based on 84 and for Case 2
  • 15. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME AREA UNDER ROC CURVE 1.2 1 AREA 0.8 0.6 GCBACS 0.4 VFF 0.2 0 0 ALGORITHM NAME Figure 8: Area under Curve considering and for Case 2 CLASSIFICATION ERROR 0.4 CLASSIFICATION ERROR GCBACS VFF 0.2 0 0 1 2 3 Figure 9: Classification Error considering 4 and 5 for Case 2 The results presented this paper prove that the proposed GCBACS is efficient and capable of analysis of crowd behaviors in the case of sparse (i.e. Case 2) and dense (i.e. Case 1) crowd surveillance videos. The GCBACS exhibits better performance when compared to the state of art Spatio-Temporal Viscous Fluid Field (VFF) method. 5. CONCLUSION AND FUTURE WORK Monitoring surveillance videos is currently achieved manually due to the random behavior of people in the crowd. Modelling and classification of crowd activities is a problem that exists. To address this issue, a computer aided Crowd Behavior Analysis and Classification System (GCBACS) is discussed in this paper. The current research systems in place do not consider inter personal behavior as an important factor for crowd behavior analysis. The GCBACS emphasis on the consideration of inter personal behavior for analysis. The behavioral features of the crowd are 85
  • 16. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME derived using the streak flows. The streak flows observed in each frame of the surveillance videos are considered as graphs. The cumulative variance is computed from the graphs and a threshold based scheme is used in to classify the crowd activity into normal and abnormal behaviors. The results presented in this paper discuss the performance evaluation between and the Spatio-Temporal Viscous Fluid Field method for sparse and dense crowd cases. The performance improvement of the GCBACS is proved in terms of recognition results, ROC curves, area under the ROC curve and classification parameters. ACKNOWLEDGEMENTS The authors of this paper would like to thank Mr Ashutosh Kumar and Mr Kashyap Dhruve for their valuable suggestion and support provided during the course of the research. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] Hang Su; Hua Yang; Shibao Zheng; Yawen Fan; Sha Wei, "The Large-Scale Crowd Behavior Perception Based on Spatio-Temporal Viscous Fluid Field," Information Forensics and Security, IEEE Transactions on , vol.8, no.10, pp.1575, 1589, Oct. 2013.doi: 10.1109/TIFS.2013.2277773. T. Zhao and R. Nevatia, “Bayesian human segmentation in crowded situations,” in Proc. IEEE Comput. Soc. Conf. Comput. Vis. PatternRecognit., vol. 2. Jun. 2003, pp. 459–466. I. Haritaoglu, D. Harwood, and L. Davis, “W4: Real-time surveillance of people and their activities,” IEEE Trans. Pattern Anal. Mach. Intell.,vol. 22, no. 8, pp. 809–830, Aug. 2000. E. Andrade and R. Fisher, “Hidden markov models for optical flow analysis in crowds, ”in Proc. 18th Int. Conf. Pattern Recognit., vol. 1. Sep. 2006, pp. 460–463. S. Ali and M. Shah, “A Lagrangian particle dynamics approach for crowd flow segmentation and stability analysis,” in Proc. IEEE Int. Conf. Com- put. Vis. Pattern Recognit., 2007, pp. 1–6. Y. Ma and P. Cisar, “Activity representation in crowd,” in Proc. Joint IAPR Int. Workshop Struct., Syntactic, Stat. Pattern Recognit., Dec. 4–6, 2008, pp. 107–116. S. Ali and M. Shah, “Floor fields for tracking in high density crowd scenes,” in Proc. Eur. Conf. Comput. Vis., 2008, pp. 1–14. Y. Yang, J. Liu, and M. Shah, “Video scene understanding using multi-scale analysis”, in Proc. IEEE Int. Conf. Comput. Vis., 2009, pp. 1669–1676. E. Andrade, R. Fisher, and S. Blunsden, “Modelling crowd scenes for event detection,” in Proc. 19th Int. Conf. Pattern Recognit., Sep. 2006, vol. 1, pp. 175–178. L. Kratz and K. Nishino, “Anomaly detection in extremely crowded scenes using spatiotemporal motion pattern models,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2009, pp. 1446–1453. R. Mehran, A. Oyama, and M. Shah, “Abnormal crowd behavior detection using social force model,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2009, pp. 935–942. J. Kim and K. Grauman, “Observe locally, infer globally: A space-time MRF for detecting abnormal activities with incremental updates,” in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit., 2009, pp. 2921–2928. K Prazdny, On the information in optical flows, Computer Vision, Graphics, and Image Processing, Volume 22, Issue 2, May 1983, Pages 239-259, ISSN 0734-189X, http://dx.doi.org/10.1016/0734-189X(83)90067-1. 86
  • 17. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 2, February (2014), pp. 71-87 © IAEME [14] Ali, S.; Shah, M., "A Lagrangian Particle Dynamics Approach for Crowd Flow Segmentation and Stability Analysis," Computer Vision and Pattern Recognition, 2007. CVPR '07. IEEE Conference on, vol., no., pp.1,6, 17-22 June 2007. doi: 10.1109/CVPR.2007.382977 [15] R. Mehran, B.E. Moore and M. Shah, & ldquo, A Streakline Representation of Flow in Crowded Scenes,&rdquo, Proc. 11th European Conf. Computer Vision (ECCV ',10), pp. 439-452, 2010 [16] X. Wang, X. Ma, and W. E. L. Grimson, “Unsupervised activity perception in crowded and complicated scenes using hierarchical bayesian models,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 3, pp. 539–555, Mar. 2008 [17] Hang Su; Hua Yang; Shibao Zheng; Yawen Fan; Sha Wei, "The Large-Scale Crowd Behavior Perception Based on Spatio-Temporal Viscous Fluid Field," Information Forensics and Security, IEEE Transactions on , vol.8, no.10, pp.1575, 1589, Oct. 2013.doi: 10.1109/TIFS.2013.2277773 [18] Si Wu, Hau-San Wong, and Zhiwen Yu," A Bayesian Model for Crowd Escape Behavior Detection", IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 24, NO. 1, JANUARY 2014 [19] Berkan Solmaz, Brian E., Mubarak Shah, "Identifying Behaviors in Crowd Scenes Using Stability Analysis for Dynamical Systems "IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 34, NO. 10, OCTOBER 2012 [20] Duan-Yu Chen and Po-Chung Huang, "Visual-Based Human Crowds Behavior Analysis Based on Graph Modeling and Matching" IEEE SENSORS JOURNAL, VOL. 13, NO. 6, JUNE 2013. [21] Nuria Pelechano and Norman I. Badler, "Modeling Crowd and Trained Leader Behavior during Building Evacuation", IEEE Computer Society 2006. [22] T. Corpetti, E. M´emin, and P. P´erez. Extraction of singular points from dense motion fields: an analytic approach. J. of Math. Im. and Vis., 19(3):175–198, 2003. [23] Unusual crowd activity dataset of University of Minnesota, available from http://mha.cs.umn.edu/movies/crowdactivity-all.avi. [24] A. Wald, "Statistical decision functions," Wiley, New York, 1950. [25] Chandramouli.H, Dr. Somashekhar C Desai, K S Jagadeesh and Kashyap D Dhruve, “Elephant Swarm Optimization for Wireless Sensor Networks –A Cross Layer Mechanism”, International Journal of Computer Engineering & Technology (IJCET), Volume 4, Issue 2, 2013, pp. 45 - 60, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375. [26] Kavita P. Mahajan and Prof. S.V.Patil, “Tracking and Counting Human in Visual Surveillance System”, International Journal of Electronics and Communication Engineering & Technology (IJECET), Volume 3, Issue 3, 2012, pp. 139 - 146, ISSN Print: 0976- 6464, ISSN Online: 0976 –6472. [27] Levina T, Dr. S C Lingareddy and Kashyap Dhruve, “Dynamic Expiration Enabled Role Based Access Control Model (Deerbac) for Cloud Computing Environment”, International Journal of Computer Engineering & Technology (IJCET), Volume 4, Issue 5, 2013, pp. 115 - 137, ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375. 87