Kernel based Object Tracking using Color Histogram Technique

Transcription

1 Kernel based Object Tracking using Color Histogra Technique 1 Prajna Pariita Dash, 2 Dipti patra, 3 Sudhansu Kuar Mishra, 4 Jagannath Sethi, 1,2 Dept of Electrical engineering, NIT, Rourkela, India 3 Departent of EEE, BIT, Mesra, India 4 Departent of AE&I, ITER Bhubaneswar pdash6@gail.co, dpatra@nitrkl.ac.in, sudhansu.nit@gail.co, jagannathsethi@gail.co Abstract. Object tracking is the process of locating oving objects in the consecutive video fraes. Real tie object tracking is a challenging proble in the field of coputer vision, otion-based recognition, autoated surveillance, traffic onitoring, augented reality, object based video copression etc. In this paper kernel based object tracking using color histogra technique has been applied for different challenging situations. Covariance tracking algorith has also been applied to the sae proble. Fro the siulation studies it is clear that this two techniques effectively handle various challenges like occlusion, illuination changes etc. Experiental results reveal that the histogra based ethod is efficient in ters of coputation tie and covariance tracker is better in ters of detection rate. Keywords: Object tracking, Color histogra, Occlusion, Covariance. 1 Introduction The object detection and tracking are the ost essential coponents of coputer vision applications ranging fro consuer electronics to sart weapons. Fast and reliable detection of oving objects is iportant for applications such as video surveillance, navigation, traffic anageent, video broadcasting, teleconferencing, huan-coputer interface etc. During video surveillance, it facilitates understanding of otion patterns to uncover suspicious events. In navigation systes it eploys to keep the vehicles in lanes and prevent collisions. In traffic anageent systes it controls the flow of vehicles to keep traffic oving soothly. Video broadcasting akes use of this to better copress data, iproving download speeds for users accessing video files on the internet. Tracking of objects becoes coplex due to several probles such as loss of inforation caused by the projection of 3D world on 2D iages, noise in iages, coplex otion of objects, 28

2 non-rigid or articulated nature of objects, partial and full object occlusion, change in scene illuination changes in background etc. Concise survey on video tracking is given by Eanuele Trucco et al. and Alper et al. in [1,2]. Researchers have proposed any ethods for object tracking. The teplates and density-based appearance odels being proposed by Schweitzer et al. in [3]. Instead of teplates, color histogras or ixture odels can be coputed by using the appearance of pixels inside the rectangular or ellipsoidal regions. Fieguth and Terzopoulos [4] generate object odels by finding the ean color of the pixels inside the rectangular object region. Coaniciu and Meer [5] use a weighted histogra coputed fro a circular region to represent the object. Coaniciu used a joint spatial-color histogra instead of just a color histogra [6]. In [7], Kang et al. used histogras of color and edges as the object odels. Alper used object tracking by asyetric kernel ean shift with autoatic scale and orientation selection [8]. Zoran in [9] applied color histogra based non-rigid object tracking algorith using ean shift kernel approach. Liu used adaptive teplate block for atching block of target in [10]. The rest of the paper is organized as follows. The Fundaentals of Object Tracking in Video Sequences areoutlined in Section 2. The two techniques for object tracking is presented and discussed in Section 3. Section 4 provides the siulation results of present studies. Finally the conclusion of the investigation is presented and further possible extension of the work is outlined in Section 5. 2 Fundaentals of Object Tracking in Video Sequences The basic flow diagra of object tracking is as follows. Video fraes Object Represen -tation Feature Selectio n Object detection Object tracking Fig.1.Block diagra of Object tracking process The object can be represented as a single point or a set of points. It ay represented as priitive geoetric shapes like rectangle, ellipse or circle. It ay be object silhouette, object contour, articulated shape odels, skeletal odel etc. The shape representation is cobined with the appearance representations for tracking purpose. Soe coonly used appearance representations are probability densities of object appearance, teplates, active appearance odel and ulti view appearance odel. Feature selection is closely related to the object representation. Soe of the coonly used features are color, edges, texture, optical flow, gradient etc. Every tracking ethod 29

3 requires an object detection echanis either in every frae or when the object first appears in the video. Soe coonly used object detection ethods are: point detectors, background subtraction, segentation and supervised learning. After detection of the object the tracker s task is to generate the trajectory of an object over tie by locating the object position in every frae of the video. 3 The Algoriths for Object Tracking 3.1 Kernel-based object tracking using color histogra In this ethod a feature space is chosen to characterize the target. The reference target odel is represented by its pdf, q in the feature space and in the subsequent frae, a candidate odel is defined at location y and is characterized by the pdf, p (y).the pdfs are estiated fro the -bin histogras. A siilar function ˆ ( y) called as the Bhattacharyya coefficient between pˆ and qˆ plays the role of likelihood and its local axia in iage indicate the presence of objects. The target ode q ˆ {ˆ }, q ˆ 1 (1) q u u 1... u 1 u The candidate odel p ˆ(y) { ˆ ( y)}, p ˆ 1 (2) p u u 1... The distance between two discrete distributions is defined as d ( y) 1 [ˆ p(y), qˆ] (3) ˆ( ˆ u 1 where [ p y), q] is the Bhattacharyya coefficient. ˆ ( y) [ˆ( p y), qˆ] pˆ ( y) ˆ (4) u 1 u q u The ain advantages of this ethods are :(i) this ethod successfully coped with coplex caera otion, partial occlusion of the target, presence of significant clutter, and large variations in target scale and appearance (ii) it is a general object detection algorith which runs very fast (ii) it is suitable for odels that have ultiple doinant colors (iii) when there is no object, global optiu needs uch less tie (iv) position, size, and orientation can be siultaneously and precisely detected and tracked. The ain disadvantage of this ethod are: (i) in this ethod the spatial inforation of the target is lost (ii) the siilarity easures like Bhattacharya coefficients and Kullback- Leibler divergence are not very discriinative (iii) representation of object histogra disregards the spatial arrangeent of the features and do not scale to higher diensions u 30

4 (iv) as it depends on the color feature, it can not give good perforance when an object and its background have siilar colors (v) the color of an object depends on illuination, view point, and caera paraeters that tend to change during a long tracking process. So fixed color features are not discriinative enough. 3.2 Object tracking using covariance tracker For a given object region, the covariance atrix of soe features is coputed. It is considered as odel of the object. In the current frae the region that has iniu covariance distance fro the odel is deterined and that region is assigned as estiated location. Let I be the observed one diensional intensity iage. F be the W H d diensional feature iage extracted fro I. F( x, y) ( I, x, y) (5) for a given rectangular window region R F, { f k } k 1... n is the d diensional feature vector inside R. f [ x, y, I( x, y), I ( x, y), I ( x, y)] (6) k x xx M N rectangular region R is calculated as follows. The covariance atrix for the 1 T CR ( fk μ R )( fk μ R ) (7) MN where μ is the vector of the ean of the corresponding features for the points within the R region R. To get the ost siilar region to the target object, the distance between the covariance atrices corresponding to the target object window and the candidate region is calculated as: 2 ( C,C ) ln ( C, C ) (8) i j k i where C, C ) are the generalized Eigen values of Ci and k ( i j j C j coputed fro C x k i k Cjxk 0, k 1,2,... d (9) where x k are the generalized Eigen vectors. At each frae we search the whole iage to find the region which has the sallest distance fro the current object odel. The best atching region deterines the location of the object in the current frae. The ain advantages of this ethod are: (i) the covariance atrix can take any possible features such as coordinate, color, gradient, edge, texture, otion etc. Hence it captures the spatial and statistical properties (ii) the covariance atrices are low- 31

5 diensional copared to other region descriptor and due to syetry C R has only ( d 2 d) 2 distinct values (iii) it works detecting single as well as ultiple rigid and non-rigid objects (iv) it works satisfactorily at object deforations and appearance change (v) noise is largely filtered out during covariance coputation (vi) the covariance atrix of any region has the sae size, thus it enables coparing any regions without being restricted to a constant window size (viii) covariance is invariant to the ean changes such as identical shifting of color values. This becae an advantageous property when objects are tracked under varying illuination conditions. The ain disadvantages of this ethod are: (i) The covariance atrices lie in a Rieannian space, hence an appropriate distance etric has to be used when coparing regions (ii) coputation is not very fast as tracking required global search. Hence the tracking of very fast oving object is difficult. 4 Siulation results Fig. 2.(a) Fig. 2.(b) Ball sequence frae 1,10,25(left to right) Fig.1(a): Histogra based, Fig 1(b): Covariance based 32

6 1/58 29/58 57/58 1/58 Fig.3.(a) 29/58 57/58 Fig. 3.(b) Table tennis player sequence frae 1,29,57(left to right) Fig.2(a): Histogra based, fig 2(b): Covariance based ethod. Fro this results it clear that all ethods are able to track the object efficiently. Fig..4(a) Fig.4.(b) Figure.6.Football player sequence2 with uneven lighting condition:frae 26,49,76 (left to right) Fig.3.(a): Histogra based, Fig.3(b): Covariance based 33

7 Detection rate is the ratio of the nuber of fraes the object location is accurately estiated to the total nuber of fraes in the sequence. The detection rate of two ethods applied to three different video sequences are shown in Table 1. The covariance tracking ethod is ore efficient as copared to other ethod in ters of detection rate.. Table.1 Coparison of detection rate of two ethods Histogra Based Frae Missed/Total Detection rate Covariance ethod Frae Missed/total Detection rate Ball video 2/ % 0/92 % Table Tennis video 2/ % 0/58 % Football video 58/ % 16/ % Reark Perforan ce of covariance ethod is best in- ters An experient has been carried out to copare the two ethods in ters of CPU tie required in second to detect object in each frae and detection rate. Table.2 Coparison of CPU tie between two ethods Histogra Based Covariance ethod CPU Tie/frae CPU Tie/frae Ball video 0 sec 600sec Tennis video 5sec 700 sec Reark Perforance of histogra based ethod is better in ters of coputation tie Football video with uneven lighting condition 580 sec 730 sec The experient results are shown in Table 2 which reveals that for a histogra based ethod, the search takes about ~0 sec/frae for ball video which is less as copared to covariance ethod. The progra is run in MATLAB with a p4, 3.2 GHz /c for 320x240 iages. 34

8 5 Conclusion and future work In this paper two techniques have been applied for object tracking for different challenging situations. Fro the experients results it can be conclude that the object tracking using this two ethodsare a siple and efficiently handle the challenging situations like occlusion, ultiple objects, faster object, coplex background, etc. The kernel based object tracking using color histogra ethod is not very good in ters of detection rate but its coputational tie is less as copared to the other ethod. The object tracking using covariance tracker ethod is ore efficient as copared to histogra ethod in ters of detection rate. Future research work in the area includes incorporation of other features like texture and otion vector to iprove the robustness of tracking algorith. Statistical odels based approach ay be applied to the proble. Evolutionary coputing ethods for object initialization ay apply to the proble. References [1] EanueleTrucco and KonstantinosPlakas, (6) Video Tracking: A Concise survey IEEE Journal.Ocean Eng., vol. 31, no. 2, pp [2] AlperYilaz, Oar Javed and Mubarak Shah,(6) Object Tracking: A Survey ACM Coput. Surv.38,4,Article 13,Deceber. [3] H.Schweitzer, J.W.Bell, andf.wu. (2), Very fast teplate atching.european Conference on Coputer Vision (ECCV).pp [4] P.Fieguth, and D.Terzopoulos.(1997) Color-based tracking of heads and other obile objects at videofrae rates. In IEEE Conference on Coputer Vision and Pattern Recognition (CVPR).pp [5] D. Coaniciu,V.Raesh, and P.Meer.(3) Kernel-based object tracking. IEEE Trans. Patt. Analy. Mach. Intell. 25, pp [6] D.Coaniciu, P. Meer, (2) Mean shift: A robust approach toward feature space analysis. IEEE Trans. Patt. Analy.Mach. Intell.24, 5, pp [7] J.Kang, I. Cohen, G. Medioni.(4) Object reacquisition using geoetric invariant appearanceodel.international Conference on Pattern Recongnition (ICPR).pp [8] AlperYilaz, (7) "Object Tracking by Asyetric Kernel Mean Shiftwith Autoatic Scale and Orientation Selection," IEEE CoputerSociety Conference on Coputer Vision and Pattern Recognition, pp.1-6. [9] Hong Yu, Jie Wei, Jin Li, (9) "Object Tracking by Mean Shift Based oncolor Distribution and Siulated Annealing," International Seinaron Future Inforation Technology and Manageent Engineering, pp [10] E. Zhou,C.Liu, Y. Sun, Z. Wang. S.Gong.(2010) Adaptive tracking window updating algorith based on particle filtering. IEEE International congress on Iage and Signal processing.pp