Object Tracking by occlusion detection via sparse learning

object tracking occlusion handling and tracking objects through occlusions using improved and object tracking robust to occlusion and size change
Dr.KeiraGibbs Profile Pic
Dr.KeiraGibbs,United States,Researcher
Published Date:11-11-2017
Your Website URL(Optional)
Comment
I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 Published Online January 2016 in MECS (http://www.mecs-press.org/) DOI: 10.5815/ijigsp.2016.01.03 Scale Adaptive Object Tracker with Occlusion Handling Ramaravind K M National Institute of Technology, Tiruchirappalli, India E-mail: raam.arvind93gmail.com Shravan T R, Omkar S N Indian Institute of Science, Bangalore, India E-mail: shravantrgmail.com, omkaraero.iisc.ernet.in Abstract—Real-time object tracking is one of the most include three steps: (a) extracting features of the object of crucial tasks in the field of computer vision. Many interest in the initial frame of a continuous video stream, different approaches have been proposed and (b) detecting object in the successive frames by finding implemented to track an object in a video sequence. One object of similar features, (c) tracking detected object possible way is to use mean shift algorithm which is from one frame to another. Mean-shift algorithm10 12, considered to be the simplest and satisfactorily efficient which comprises all of the above mentioned steps, is used method to track objects despite few drawbacks. This as object tracker for the rest of the discussion due to its paper proposes a different approach to solving two simplicity and efficiency for implementation in real-time typical issues existing in tracking algorithms like mean tracking 1 2 3. Mean shift algorithm is a shift: (1) adaptively estimating the scale of the object and nonparametric density estimator which iteratively (2) handling occlusions. The log likelihood function is computes the nearest mode of a sample distribution 4 used to extract object pixels and estimate the scale of the 11. It considers that the input points are sampled from object. The Extreme learning machine is applied to train an underlying probability density function. It starts with a the radial basis function neural network to search for the search window that is positioned over these sample points object in case of occlusion or local convergence of mean which continuously moves till the centre of this window shift. The experimental results show that the proposed coincides with the maxima of that distribution. algorithm can handle occlusion and estimate object scale But when the object is scaled in any sense with respect effectively with less computational load making it to the camera view, classical mean-shift still using the suitable for real-time implementation. previous window size fails to correctly identify the object. Unlike CAMSHIFT 5 which uses moment of weight images determined by the target model and SOAMST 6 Index Terms—Object Tracking, Mean-shift, RBF Neural using moment of weight images determined by both Networks, Scale estimation, Occlusion handling. target and target candidate model to determine the scale changes, the proposed algorithm uses log-likelihood function to extract only the objet pixels from the I. INTRODUCTION infelicitous sized search window and thereby estimating Due to its ubiquitous applications such as automated the scale of the object and updating the search window surveillance, traffic monitoring, human identification, size. But this procedure can be applied only when most of human-computer interactions, vehicle navigation, the object pixels are found inside the search window or in aerospace applications etc, object tracking is considered other words the mean-shift should be converged. as a cardinal task in the field of computer vision 1. Again mean-shift fails when the object is taken out of Despite the advancement of technology, several frame or occluded by any other object during tracking. challenges are encountered in real time visual object Thus handling occlusions and estimating the scale of the tracking 1 2, such as change in illumination, partial or object when the mean-shift is not converged becomes a full object occlusions, complex object shapes, noise in matter of concern. This paper proposes the use of radial images, scale change, similar objects, cluttered basis function neural network for solving these issues. background, real-time processing requirements etc Applying Extreme learning machine algorithm 7 to making the design of a perfect algorithm virtually RBF neural networks 8 9 has been proved to provide impossible. Evidently constraints, depending on the good generalization performance in most cases and learn application, need to be maintained in developing object faster than conventional algorithms. Thus, several search tracking algorithms solving some challenges while regions are formed wisely depending on the application compromising some. and the trained RBF NN is used to identify the presence People have come up with numerous visual object of object pixels in these regions. In this paper, only RGB tracking algorithms 1 3. Few of them primarily values of the object pixels are used as features while this Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 28 Scale Adaptive Object Tracker with Occlusion Handling selection actually depends on the computational feature of all the pixels in the rectangular window is complexity and accuracy required by the application. The extracted. Feature extraction is the one of the most proposed solution incurs reduced computational load and computationally expensive step in object tracking. is promising for real-time applications. Features represent the object and evidently extracting The remainder of the paper is organized as follows: more features will more accurately represent the target Section II describes the overview of proposed tracking object for tracking. But cognizant of the fact that algorithm (illustrated in Fig.1). Section III describes extracting more features will impart computational load radial basis function network for handling occlusions and and the prime motive of the proposed algorithm is to improving efficiency. The experimental results and handle occlusion and being robust to scale change in real comparisons are explained in Section IV and discussions time, pixel based colour features is used in the form of R- and conclusions are given in section V. G-B paradigm. Now we divide each channel of RGB colour space into k bins or intervals, each constituting an element of the feature space. So we get a feature space vector of bins II. SYSTEM OVERVIEW or elements. Then we find the correspondence of every pixel to the bins so constructed and prepare a histogram. B. Background weighted target and target candidate model The colour histogram formed using all the pixels within the rectangular window is used to represent the target. The target model 10 11 and 12 is given as: ̂ ̂ + (1) ∑ (‖ ‖ ) , ( ) - ̂ (2) Where ̂ is the target model, ̂ is probability of element of ̂ , is the Krnocker delta function, ( ) associates the pixels to histogram bin, is an isotropic kernel profile, and constant ∑ (‖ ‖ ) = . The target model always consists of background features. When the target and background are more similar, the localization accuracy of object will be decreased and finding new target location becomes difficult 11. To reduce the interference of salient background features in target modelling, a representation model of background features is separately built and the target model is updated based on the background model 11 12. The background is represented as ̂+ with ∑ ̂ in 11 12 and is calculated by surrounding area of target with appropriate window size Denoted by ̂ is the minimal non-zero value in ̂ + ̂ . The coefficients . / are ̂ used to define a transformation between target model and target candidate model which reduces the weights of those features with low i.e., the salient features in the background 11. Thus the new target model becomes , - ̂ ∑ (‖ ‖ ) ( ) (3) Fig.1. Process Flow of the Proposed Method Where constant ∑ (‖ ‖ ) , ( ) - A. Object selection and Features extraction The target candidate model is constructed in a similar way as that of target model but with increased window The object to be tracked is selected manually in the size as follows first frame by specifying a rectangular window. Then the Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 Scale Adaptive Object Tracker with Occlusion Handling 29 ( ) + ̂ ̂( ) (4) a small non-zero value to avoid numerical instability. The log-likelihood outputs positive values for object pixels and negative values for background pixels. Any ̂( ) ∑ (‖ ‖ ) , ( ) - (5) pixel is declared as object pixel if T. For most of our experiments, T was kept constant at 0.8 which gave Where ̂ ( ) is the target candidate model, ̂( ) is satisfactory results. This is further improved by applying probability of element of ̂ ( ), + are the morphological operations on the obtained likelihood to accurately represent the object pixels and eliminate pixels in the target candidate region centred at y, is the outliers. Fig.3 offers an overview of the process. bandwidth, and the normalized constant ∑ = (‖ ‖ ) . Here the target model alone is updated based on background information and not the target candidate model according to 11. C. Extracting object pixels The background weighted target model reduces the weights of salient background features and helps in reducing the background interference in target model. This target model along with background histogram is used to separate the object pixels form background pixel. The log-likelihood ratio of object-background region is used to determine the object pixels as done in 8. Fig.2 describes the process. Fig.3. Object and Non-Object Pixel Extraction D. Adaptive scale estimation of objects The classical mean-shift approach does not estimate (a) (b) the scale of object being tracked neither it adaptively changes the size of search window. In the proposed algorithm, log-likelihood is used not only to extract object pixels but also to address the scaling issues in mean-shift. After the separation of object and background pixels, we get a rough contour of the object. This outer contour is used for estimating the scale of the object and (c) (d) searching window size (as illustrated in Fig.3 and Fig.4). (e) (f) Fig.2 (a), Fig.2 (b) Shows the Actual Tracked Images. Fig.2 (c), Fig.2 (d) (a) (b) Shows the Likelihood Maps ―L‖. Fig.2 (e), Fig.2 (f) Shows the Morphologically Transformed Images. For every pixel in the background window, the log- likelihood ratio is obtained as ( ) + . / (6) ( ) (c) (d) ( ) ( ) Fig.4 (a-d). Shows the Adaptive Scale Estimations in Frames where and are the probability of pixel 10,25,42,68 Respectively. belonging to object and background respectively and is Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 30 Scale Adaptive Object Tracker with Occlusion Handling E. Mean shift Tracking When current window size is equal to the initial window size, the proposed similarity coefficient is just equal to the The new target candidate location from the current Bhattacharya coefficient. location is given by mean-shift tracker 11 as When drop below a particular threshold, the algorithm is presumed to be not converged since the ∑ (‖ ‖ ) target similarity to candidate is reduced as indicated by ( ) (7) the similarity coefficient. In our experiments, this ∑ (‖ ‖ ) threshold was kept at 0.3 ̂ ∑ √ , ( ) - (8) ̂ ( ) III. OCCLUSION HANDLING AND EFFICIENCY IMPROVEMENT USING RBF NN Where ( ) is the shadow of kernel profile ( ) : ( )= ( ). The mean-shift tracker basically does the so called ―frame to frame‖ tracking of objects. So it may converge F. Background and target updating to a local mode or may not converge at all when the The target and the background colour model are object is taken out of frame or occluded in between. It is also prone to local convergence when the object is utterly initialized in the first frame of tracking. However, the background may often change during tracking due to indistinguishable to the background. The disparate modelling of the target depends heavily on feature variations of illuminations, viewpoint, occlusion etc. So there arises a need of updating the background model selection. Since we are using only RGB features to + represent our target, due to various constrained 11. First the background features ̂ and ̂ + in the current frame are calculated. Then the mentioned earlier, we seek the aid of artificial neural network to scout the object in case of occlusion and local Bhattacharya similarity between old and current convergence. background model is computed by We initially train the network with different scale and perspective of an object and then use this trained network ∑ √ ̂ ̂ (9) to search our object when they are abandoned by mean shift tracker. Fig.5 illustrates the training process. If is smaller than a particular threshold, the background model is updated and the target model is also updated based on changes in the background. In the experimental results shown in section IV, the threshold was kept at 0.5. This updating of both target and background model is quite intuitive and helps in improving the efficiency. G. Similarity Coefficient The Bhattacharya coefficient is usually computed, in tracking algorithms like mean-shift for measuring similarity between object and background model, as ∑ √ ̂ ̂ . The target candidate model is constructed from the most recent object centre and an increased version of the updated search window. This increment remains constant throughout the tracking process making the background information interfere more in the target candidate model when the object is scaled down in size. Hence Bhattacharya coefficient diminishes in magnitude when the object is scaled down, not indicating the actual similarity between target and target candidate. So we suggest a modified similarity coefficient based on updated scale of the object as: (10) Fig.5. Training RBF Neural Network Before the usage of RBF NN for our application is where is the similarity coefficient and is elucidated, it is reasonable to discuss the region in which Bhattacharya coefficient between the target and candidate the object is to be searched. When the object is occluded in the ( ) frame, and are window sizes in or gone out of frame or not converged, the possible and initial frame respectively. The denominator in locations at which the object can be seen again are here called as interest regions. The interest region may vary similarity coefficient functions as the scaling factor. Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 Scale Adaptive Object Tracker with Occlusion Handling 31 depending on the application, tracking algorithm, the size score for the associated class and the input is assigned to of object and computational efficiency required. One the node with highest score. The scores are computed by approach is to divide the entire frame into required taking a weighted sum of the activation values from every number of patches and search for the object in those RBF neuron as follows patches. The patches may or may not be overlapping. Searching object in a greater number of patches makes .‖ ‖ / ( ) ( ) ∑ the algorithm computationally expensive. In our experiments with mean shift tracker, we use appreciable (12) number of patches satisfying both accuracy and time constraint which is explained in section IV. In matrix form, where the dimension of Y, A. Radial Basis Function Neural Network and H are respectively. These output weights are calculated by Moore-Penrose RBF classifier with single hidden layer neurons generalised pseudo-inverse as follows: where incorporating fast learning algorithm 7 8 is proved to specified pseudo inverse. approximate any continuous function to desirable The output weights are calculated for input points from accuracy by overcoming many issues such as stopping each frame till the training phase gets over. At each frame criterion, learning rate, number of epochs and local we also calculated the area of objects which is nothing minima. but the total number of object pixels. Then we compute The first layer of the network consists of the inputs the area weighted output weights which are presumed which are feature vectors of sample points. We allot first to be our finalized parameters for testing phase. frames after object selection for the training phase. The exact number of frames required for training phase ∑ ( ) depends on the accuracy required. (13) ∑ The extracted RGB values of object and background pixels in the previous stages constitute the input signals where is the area of object and is the output weights for our network. So input is dimensional feature calculated in the . vector where 3 denote the RGB features and is the When the object is occluded or mean-shift is not number of pixels. The label of each pixel either as object converged by any means, we form interest regions and at or background forms the output vector which is each frame, we search the object in one or few of the dimensional vector. interest regions presuming that the object movement is The next is the hidden layer which uses Gaussian not so drastic from one frame to another, until the mean functions and is non-linear. Each hidden neuron is shift is converged. represented by a centre vector which is one of the vectors By searching, we mean that we get features from the from the training set and measure of the width of current interest region(s) and find object pixels using Gaussian function. The centres can be chosen randomly RBF NN. If the network doesn‘t return any object pixels, from the training set. Here we apply K-means clustering we shift our interest region(s) until we find the object. on the training set, once for each class and use the cluster centres as centres of hidden neurons according to 13 14. Apparently more hidden neurons provide more IV. EXPERIMENTAL RESULTS accurate results but require more computation. K-means is one of the popular and straightforward approaches and The algorithm is tested on certain video sequences of is not discussed in detail in this paper. After calculating 320240 pixels resolution on MATLAB 2012 the centres, the output of each RBF neuron is given as environment. A detailed result for ‗blob sequence‘ consisting of 334 frames is given below. The object is the .‖ ‖ / dark green coloured circular blob which is manually ( ) , (11) moved erratically by attaching it to the tip of a rod. In the first frame, the object ‗blob‘ is selected by where is the centre of hidden neuron got from k- specifying a rectangular window (green rectangle). Then means clustering and and which relate to widths of the RGB features of every object and background pixels Gaussian curve is calculated as follows 13 14: are extracted and respective histograms are computed as explained in section II. Here the RGB feature space is ∑ ‖ ‖ quantized to 16x16x16 bins ( ) and Epanechnikov kernel is used as kernel function ( ) according to 12. These histograms in Fig.6 and Fig.7 are the RGB based and object and background model respectively Then the log-likelihood function between object and , + input data set. background model is used to estimate the scale changes of the object and the object is tracked by mean shift. The The output of network consists of two nodes one for background model is updated whenever there were obvious changes in the background according to 11. object and other for non-object. Each node computes a Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 32 Scale Adaptive Object Tracker with Occlusion Handling This locates the object more accurately. Some of the frames, where target updating was done, are shown in Fig.8(a-c). During tracking, when the object is occluded or similarity coefficient drops down below a threshold, object pixels are searched in the interest regions. For this experiment, the frame is partitioned into five rectangular areas, one within each of four quadrants and one at the centre and theses areas form the necessary interest regions. Fig.9 shows one such classification by RBF NN into object and non-object at one of the interest regions. Initially, 40 frames are allocated for training for this experiment and the training accuracy of the RBF NN is given in Fig.10. Fig.9. Classification Results of Radial Basis Function Network. Each one of the Three Axes Scales from 0-255 which indicates the Intensity Values of Red, Green and Blue Channels Respectively. Fig.6. Object Histogram Fig.10. The % Training Accuracy of RBF NN. The proposed algorithm is then compared with Corrected Background weighted histogram mean shift 11 and Scale and orientation adaptive mean shift 6 algorithms. In 11, a corrected Background Weighted histogram algorithm is proposed where only the target region and not the target candidate region is updated based on background information. Whereas in 6, the order moment and the Bhattacharya coefficient between target and target candidate model is used to Fig.7. Background histogram estimate the scale and orientation changes of the target. The tracking result of proposed approach and corrected background weighted histogram (CBWH) mean-shift is first showed in Fig.8. Although CBWH does not promise to handle occlusion, the following comparison was carried out to provide a better insight of the proposed method against the most similar CBWH algorithm for object tracking. Object detected by the proposed (a) (b) (c) algorithm and CBWH is shown in green and blue Fig.8. Background Updating At Frames 53(a), 195(b) and 250(c) rectangles respectively. When the object is scaled down, the proposed approach identifies the object with correct scale estimation while in CBWH the scale of the search window is unmodified as shown in Fig.11-b. The scale of the object is correctly predicted even when the object is only partially visible as shown in Fig.11-e. When the Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 Scale Adaptive Object Tracker with Occlusion Handling 33 object is occluded in frame 127 (Fig.11-f), the proposed algorithm rightly reveals the absence of object but CBWH converges locally to the previously found object location. Subsequently the object was found in frame 132 (Fig.11-h) by the proposed algorithm although it missed (Fig.11-g) in frame 130 when the object was first restored in the view. But CBWH identifies the object in frame 195 (a) (b) (c) (Fig.11-i) when the object was brought closer to the locally converged region (d) (e) (f) Fig.12 (a-f). Tracking Results of Proposed Algorithm Vs SOAMST for Frames 1, 15 (Object Scaled Down), 25 (Object Scaled Up), 85 (a) (b) (c) (Background Change), 100 (Object Moved Fast) and 265(Local Convergence of SOAMST). Apart from the above comparisons, the number of iterations taken by the proposed algorithm and SOAMST are compared below for better clarity (Table 1). It can be noticed that from frame 1 to 100, during which the object is clearly observable and scaled at some instances, the (d) (e) (f) proposed algorithm tracks the object with a minimum number of iterations than SOAMST. Thus, it incurs less computational load than SOAMST to adaptively estimate the scale of object making it more favourable for real- time applications. After frame, the object remained occluded a number of times and for every such occurrence the iterations number reach a maximum for (g) (h) (i) the proposed method while it is unpredictable for Fig.11 (a-i). Tracking Results of Proposed Algorithm Vs CBWH Mean SOAMST as shown in Fig.13 and Fig.14 respectively due Shift for Frames 2, 14 (Object Scaled), 124, 125, 126, 127(Occlusion to local convergence. Encountered) 130(Object Missed) 132(Object Detected Again By Proposed Algorithm), 195(Object Detected By CBWH). Table 1. Mean and Standard Deviation (of Iterations) of Proposed Method and SOAMST are Tabulated for (1) Entire Sequence, (2) The next comparison was carried out between the Frames 0-100 When the Object is Scaled and Observable and (3) proposed approach and the SOAMST 6. Object detected Frames 100-334 When the Object Remains Occluded in Some Frames. by the proposed algorithm and the SOAMST is shown Frame Range SATOH SOAMST within the green rectangle and red ellipse respectively. Mean STD Mean STD The green ellipse just represents the background window 0-100 3.1900 1.0512 5.1400 2.5585 100-334 2.7094 0.9941 4.3419 3.3601 of the object, according to SOAMST. In 6, the scale and 0-334 2.8533 1.0337 4.5808 3.0996 orientation changes of the target were successfully Note: SATOH- Scale Adaptive object Tracker with Occlusion Handling. estimated. However, the algorithm worked slowly compared to the proposed one since the former involves heavy computation calculating moments of weight images. When the object is scaled up or down or changed in orientation, it is well tracked by both SOAMST and proposed algorithm as shown in Fig.12 (a-c). But when the object is moved fast in frame 100 (Fig.12-e) or its background is abruptly changed in frame 85(Fig.12-d), SOAMST was unable to track the object while the proposed algorithm successfully did in most of the conditions. SOAMST remained locally converged till the last frame (Fig.12-f). Fig.13. Number of Iterations by the Proposed Method SATOH Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 34 Scale Adaptive Object Tracker with Occlusion Handling adaptive scale estimation by extracting object pixels and finding abandoned object by radial basis function network incurs less computation burden and hence is more suitable for real-time implementation The convergence decision in the proposed approach using mean-shift object tracker is based on two parameters-similarity coefficient and restraint on the maximum number of iterations. The similarity coefficient depends on the object to be tracked. If the object is much similar to the background, the decision on similarity coefficient becomes difficult hence local convergence becomes an issue. The maximum number of iterations is Fig.14. Number of Iterations by SOAMST again restrained depending on computation constraint. This error can be minimised if more features like texture The next comparison is carried out between the and shape are used. The decision on search area or similarity coefficient as proposed in the paper and the interest regions can be tough as even the entire frame can Bhattacharya similarity used in CBWH. The similarity be searched for objects. But due to the computational coefficient, correctly predicts the similarity between the complexity we restrict ourselves to RGB feature space target and target candidate when the object is occluded, and limited interest regions. as nil (as illustrated in Fig.15). But the Bhattacharya The proposed method becomes non-parametric by coefficient which is used for similarity measurement in combining neural net with mean-shift and some amount 11 predicts the similarity incorrectly in such situations. of time must be sacrificed for the training phase and Further when the object is scaled down in frame 16, the selecting felicitous similarity coefficient for better similarity coefficient implies the actual similarity while performance. The algorithm can be improved by the Bhattacharya coefficient drops down in magnitude efficiently including textures or shapes or other robust due to scale change (as illustrated in Fig.16). features without significantly affecting the efficiency of algorithm to avoid local convergence and intelligently selecting the search region depending on the anticipated trajectory of the object which are currently under our study. ACKNOWLEDGEMENT We gratefully acknowledge the support from Indian Institute of Science for Scientific Research without which the work could not have been completed. In developing the ideas presented here, we also received helpful inputs on neural networks from J. Senthilnath. Fig.15. Similarity Coefficient between Target and Target Candidate as REFERENCES Done in SATOH 1 Yilmaz A., Javed O., and Shah M. ―Object Tracking: a Survey‖, ACM Computing Surveys, 2006, 38, (4), Article 13, pp.2-5. 2 NM Artner, ―A comparison of mean shift tracking th methods‖, 12 Central European Seminar on computer graphics, 2008, pp.1-3. 3 B Karasulu, ―Review and evaluation of well-known methods for moving object detection and tracking‖, Journal of aeronautics and space technology, 2010, pp.1-4, 6-9. 4 Fukunaga F. and Hostetler L. D. ―The Estimation of the Gradient of a Density Function, with Applications in Pattern Recognition‖, IEEE Trans. on Information Theory, 1975, 21, (1), pp.32-40. 5 Bradski G. ―Computer Vision Face Tracking for Use in a Fig.16. Bhattacharya Similarity between Target and Target Candidate as Perceptual User Interface‖, Intel Technology Journal, Done in SOAMST 1998, pp.4-7. 6 J Ning, L Zhang, D Zhang, C Wu, ―Scale and orientation adaptive mean shift tracking‖, IET Computer Vision, 2012, pp.6-9. V. DISCUSSIONS AND CONCLUSIONS 7 G.-B. Huang, Q. Y. Zhu, and C. K. Siew, ―Extreme In this paper, we proposed a hybrid solution to some of learning machine: Theory and applications,‖ the common problems existing in object tracking. The Neurocomputing, 2006, pp.3-9. Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35 Scale Adaptive Object Tracker with Occlusion Handling 35 8 RV Babu, S Suresh, A Makur, ―Robust object tracking Authors’ Profiles with radial basis function networks‖, ICASS, IEEE international Conference, 2007, pp.2-3. Mr. Ramaravind K M is a recent 9 G.-B. Huang and C.-K. Siew, ―Extreme learning machine undergraduate student researcher from National with randomly assigned rbf kernels,‖ International Journal Institute of Technology, Tiruchirappalli where of Information Technology, vol. 11, no. 1, pp. 16–24, 200. he pursued his Bachelors in Instrumentation 10 Comaniciu D., Ramesh V., and Meer P. ―Real-Time and Control Engineering. His research interests Tracking of Non-Rigid Objects Using Mean Shift‖. Proc. include Artificial Intelligence, Computer Vision, IEEE Conf. Computer Vision and Pattern Recognition, Machine Learning and Robotics. Hilton Head, SC, USA, June, 2000, pp. 142-149. 11 J Ning, L Zhang , C Wu, ―Robust mean-shift tracking with corrected background weighted histogram‖, IET Mr. Shravan T R is currently a Master Student Computer Vision, 2012, pp.4-6, 8-11. at TU-Delft. His specialisation is in the field of 12 Comaniciu D., Ramesh V. and Meer P. ―Kernel-Based Bio-Robotics. His research interests include Object Tracking‖, IEEE Trans. Pattern, Anal. Machine Human-Robot systems, Image Processing, Intell., 2003, 25, (2), pp. 564-577. Neural Networks. 13 F Schwenker, HA Kestler, G Palm , ―Three learning phases for radial basis function networks‖, Neural network, 2001, pp.4-8. 14 ―Radial Basis Function Network (RBFN) tutorial‖, Dr S N Omkar is the principal research http://chrisjmccormick.wordpress.com/2013/08/15/radial- scientist at Indian Institute of Science, basis-function-network-rbfn-tutorial/, accessed Aug 2013. Bangalore. He is currently working in the field of neural networks, unmanned aerial vehicles, bio-mechanics, remote sensing and Geographic information system. How to cite this paper: Ramaravind K M, Shravan T R, Omkar S N,"Scale Adaptive Object Tracker with Occlusion Handling", International Journal of Image, Graphics and Signal Processing(IJIGSP), Vol.8, No.1, pp.27-35, 2016.DOI: 10.5815/ijigsp.2016.01.03 Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 1, 27-35