Efficient and Accurate Object Tracking for Mobile Mapping System Using GIS Object Templates Lingeswaran S Segmentation

and Tracking ofVGIS Object for MMS

Abstract: Segmentation and tracking of objects in a 2D image
sequence is an important and challenging field of wide usages including change detection, object-based video post production, content-based indexing and retrieval, surveillance. This paper proposes a novel hybrid method to segment the GIS object in DMI sequences using pyramid decomposition. The proposed method involves image enhancement strategy to hypothesis for foreground background identification and applying watershed transformation to achieve foreground background segmentation more effectively. The proposed method also uses image enhancement and denoising strategy to enhance the frameset to properly track the Object of interest in the given spatiotemporal data set using Gaussian filter. Pyramid decomposition based strategy is used to track the GIS object in sequential frame which involves DMI. Pixels having the highest gradient magnitude intensities (GMIs) correspond to watershed lines, which represent the region boundaries are tracked in subsequent frames. The object template is updated with the frame set to increase the tracking performance.

Keywords-- Digital Measurable Image (DMI), Segmentation,

Object, Color Histogram Back-projection, Watershed, Denoising

I. INTRODUCTION

n quest of low cost, compact size, and efficient tracking methodology based Mobile Mapping system has always been a keen interest of the automobile operators. The tracking of object of interest (OOI) in the field of digital Measureable Image has attracted lots of researchers. DMI sequence acquired by a mobile imaging platform with a much larger time gap; they are not competent to get spatiotemporal segmentation of GIS object in DMI sequence. Automatic GIS data extraction and update is still an unsolved puzzle and the bottle- neck of the MMS [1], [10]. Object-based image sequence segmentation techniques can be grouped into different categories, for instance, supervised, unsupervised or semi-supervised; region based or boundary based; high level or low level; local information based or global information based; segmentation of mobile objects or static objects. DMI sequence acquired by a mobile imaging platform with a much larger time gap; they are not competent to get spatiotemporal segmentation of GIS object in DMI sequence.. Most arresting information in MMS is about GIS objects, e.g. guideposts, central lines and so on. Accurate and efficient segmentation of GIS objects in DMI sequences is a necessary preprocessing for the MMS and there is still a lack of credible work on this issue. Several types of active contour have been utilized for video object segmentation and tracking. Vaswani et al

proposed that particle filter combined with level set based active contour can be used to segment moving and deformable object in an image sequence [15]. Graph or hyper graph is used in some region based segmentation method.SVM based strategy have widely used for clustering in hyper graph lane. Some region based methods employ clustering operation or region splitting and growing in the feature space, which is usually formed by motion vectors, spatial features or appearance features like color, texture, and position. The proposed approach uses watershed algorithm for the foreground and background segmentation process. Watershed approach has been introduced in semi automatic video object racking and shown improvised result from the previously proposed method. Patch color histogram approach have been used for the exact tracking of the object from the various frame of the video sequences. In order to consider the local information of the provided video sequence, image enhancement including Gaussian filter have been employed to enhance the quality output. This paper is organized as follows: Section II proposes Watershed Transformation for image segmentation. Section III proposes adaptive local filters while the experimental result is shown in section IV and conclusion is defined in section V.

II. WATERSHED TRANSFORMATION FOR IMAGE

SEGMENTATION Image segmentation is one of the widely utilized tools for the analysis and extraction of the object of interest from the provided images. Watershed transformation is one of the powerful tools for image segmentation and overcomes the problem of disconnected contours and false edges. Watershed transformation can be utilized for reducing the conflict of false edge definition and enhanced tracking. It considers the gradient magnitude of an image or a distance transformation of a binary image as a topographic surface. Pixels having the highest gradient magnitude intensities (GMIs) correspond to watershed lines, which represent the region boundaries. Water placed on any pixel enclosed by a common watershed line flows downhill to a common local intensity minimum (LIM). Pixels draining to a common minimum form a catch basin, which represents a segment. However, in practice, this transform produces severe over- segmentation due to noise or local irregularities in the gradient image. In order to cope with the problem of noise and local irregularities image enhancement strategy have

been proposed in the section III. Enhancement strategy includes adaptive Gaussian filter to increase the performance of the available video sequence. A previously defined set of markers can be used as a method to enhance the watershed transformation segmentation results.

effective for processing. The Frame set is denoised using the Gaussian filter to increase the range of prediction from the given OOI.

A.

Hypothesis For Figure Background Identification And Auto-Marked Watershed Transformation According to characteristics of GIS OOI (Geographical information based object of interest), we propose the following hypothesis for foreground background identification of GIS OOI template. Pixels having gradient magnitude intensities near to the region boundary belong to the region only;

While in the object template, pixels near the bounding box (boundary) belong to background; the template object is always less in size than the given GIS video sequences; The OOI lies in the center area of the template image.

Although the foreground-background segmentation result is coarse, most of the background can be eliminated and the remainder foreground is sufficient to compute the object of interest from the given frames. The foregroundbackground segmentation enhances the performance of segmentation process and increase the quality of the segmented region. Thus the region from the given frame set is obtained for processing and tracked. However the region so tracked cannot be exactly revealed with the OOI. Patch color histogram strategy is being used for the exact segmentation of the tracked region.

Fig: 2 Local Statistics Estimation We assume that only background objects are present in the first image plane, so that spatial clustering of the first frame gives us the distribution of regions that characterize the scene. Foreground distributions are created in the future image planes as the algorithm processes. The model is initialized using the spatial segmentation of the first frame and the parameters of the distributions are initialized using the local statistics calculation shown in Figure 1. If the number of pixels that belong to region i within the neighborhood is too small, the estimates of i,s,t and i,s,t are not very reliable and therefore no distribution is created for this region. In this way, outliers can be discarded and distribution statistics can be more accurately estimated. Also, in the case of occlusion, background pixel distributions are preserved.

Fig1: Background and Foreground Segmentation Using Watershed Transformation

According to the above three hypothesis, we proposed an auto-marked watershed transform scheme for foreground background identification scheme.

Segment t h e GI S obj e ct te mpl at e u s i ng wa t e r s he d transformation. All the segments adjacent to boundary of the template will be abandoned. The remainder segments form a mask indicating the exact OOI. This cause the template to be more precise and

B.

Fast Patch-based Color Histogram Backprojection In this paper we propose an approximate but quick method, fast patch-based color histogram back-projection method. It is less time-consuming and more robust to object deformation.

The improvements compared to former patch-based color histogram back-projection are: Both the object template and the input image are resized to half of their initial size, for color histogram is nearly invariant to scale change. You can resize the inputs much smaller for speed if you like. The sliding step is set to one third of the object size in corresponding directions rather than 1 pixel. For most of GIS objects template are much bigger than 3 pixels, hence a higher speed will be achieved.

function to identify potential interest points that are invariant to scale and orientation. 2. Keypoint localization: At each candidate location, a detailed model is t to determine location and scale. Key points are selected based on measures of their stability. 3. Orientation assignment: One or more orientations are assigned to each key point location based on local image gradient directions. All future operations are performed on image data that has been transformed relative to the assigned orientation, scale, and location for each feature, thereby providing invariance to these transformations. 4. Keypoint descriptor: The local image gradients are measured at the selected scale in the region around each Keypoint. These are transformed into a representation that allows for signicant levels of local shape distortion and change in illumination.

For comparing two histograms, we choose the histogram intersection, shown in (2), as the correspondence measurement for efficiency consideration. Control variables: K , , W , Initialization: Obtain initial segmentation x using ACA Calculate local statistics i,s,t , i,s,t Initialize Gaussian pdfs pi using i,s,t , i,s,t Initialize foreground Gaussian pdf p f to NULL while new data yt do // Temporal Labeling If> t hreshold then pixel = foreground pixel label = K + 1 if tem plate > threshold then pixel = foreground pixel label = K + 1 else if template < threshold pixel = background pixel label = i end if // Update Calculate local statistics s,t using xt Update temporal model by applying Smoothing Smooth xt by MRF until convergence using end while The cost of extracting the scale invariant features is minimized by taking a cascade ltering approach, in which the more expensive operations are applied only at locations that pass an initial test. Following are the major stages of computation used to generate the set of image features: 1. Scale-space extrema detection: The rst stage of computation searches over all scales and image locations. It is implemented efficiently by using a difference-of-Gaussian

Fig 3: Scale Invariant Feature Location Identification

This approach has been named the Scale Invariant Feature Transform (SIFT), as it transforms image data into scaleinvariant coordinates relative to local features.

III. ENHANCEMENT FOR IMAGE EXTRACTION

The steps of our enhancement technique are as following: 1-Unsharp masking step: Enhances small structures and bring out the hidden details in the image by using unsharp masking. It only sharpens the areas, which have edges or lots of details. Unsharp masking performed by generating a blurred copy of the original image by using laplacian filter, subtracting it from the original image.

I(i,j) = I0(i,j) + Ib(i,j) Where I (i,j) unsharp masking image I0(i,j) original image Ib(i,j) blurred copy

verify the proposed spatiotemporal segmentation method and performed on computer with Intel Pentium 4 2.4GHz CPU and 1GB RAM.

Multiply the unsharp masking image by a fractional value, and adding it to the original image to get the image that will be contrasted. In this step, the large features are not changed by much, but the small ones are enhanced. The result is a sharper, more detailed image. G(i,j) = I0(i,j) +k Ib(i,j) Where: g (i, j) is output image, k is scaling constant. Logical values for k vary between 0.2 and 0.7. Recently there was an attempt to perform the sharpening by local analysis of gradients. 2-Contrast enhancement step: For a grey scale images A sliding 3x3 map window moves from the left side to the right side of original image horizontally in steps starting from the images upper right corner. A pixel value in the enhanced widow dependents only on its value thats mean if the interest pixel exceeds a certain value (threshold) its value remain unchanged if the value of the pixel is under the threshold then it will be remapped .The process can be described with the mapping function O =M (i), where O and i are the new and old pixel values, respectively. M=
Fig 4: interface for extracting and matching the template

c i * i 1+ e

According to above mapping function the new value of corresponding pixel will be: O= 3
Fig 5: Interface detecting the template in frame accurately

Where c is a contrast factor determines the degree of the needed contrast. After map window reaches the right side, it returns to the left side and moves down a step. The process is repeated until the sliding window reaches the right-bottom corner of the image. For color images, before applying the slider map window the pixels with the lowest values map to 0 and the highest values to 255 other pixels value calculated based on a weighted average of the RGB. IV. EXPERIMENTAL RESULT We tested the performance of the proposed technique on grey scale and color real remote sensing images. The effectiveness of the proposed method is demonstrated via the experiment results using DMI sequences. Experiments are implemented on MATLAB R2010a to

Fig 6: Interface tracking template in the DMI Sequence

REFRENCES

[1] Peng [2] J.

Li, Cheng Wang, Hanyun Wang, Shengyong Hao, Spatiotemporal egmentation of GIS Object for Mobile Mapping System IEEE, Conference 2011. G. Allen, R. Y. D. Xu, and J. S. Jin, Object Tracking Using CamShift Algorithm and Multiple Quantized Feature Spaces, PanSydney Area Workshop on Visual Information Processin, 2003. [3] N. Apostoloff, and A. Fitzgibbon, Automatic video segmentation using spatiotemporal T-junctions, BMVC, 2006. [4] S. W. Babacan, and T. N. Pappas, Spatiotemporal algorithm for joint video segmentation and foreground detection, EUSIPCO, 2006. [5] E. Borenstein, and J. Malik, Shape Guided Object Segmentation, CVPR, 2006. [6] G. Bradski, and A. Kaehler, Learning OpenCV. OReilly Media Inc., Sebastopol, pp.194-221, 2008. [7] W. Brendel, and S. Todorovic, Video Object Segmentation by TrackingRegions ICCV, 2009.

[8] P.

L. Correia, and F. Pereira, Classification of Video Segmentation Application Scenarios, IEEE Trans. on Circuits and Systems for Video Technology, 14(5), pp. 735-741, 2004.

[9] R. [10]

Ahmed, G. C. Karmakar, and L. S. Dooley, Probabilistic Spatio- Temporal Video Object Segmentation Incorporating Shape Information, ICASSP, 2005. S. beucher, The watershed transformation applied to image segmentation, 10th Pfefferkorn Conf. on Signal and Image Processing in Microscopy and Microanalysis, 16-19 sept. 1991, Cambridge, UK, Scanning Microscopy International, suppl. 6. pp. 299-314, 1992.

[11]

J. Huang, S. Ravikumar, M. Mitra, W.J. Zhu, and R. Zabih, spatial Color Indexing and Applications, International Journal of Computer Vision, 35(3), pp. 245268, 1999. [12] Y. C. Huang, Q.S. Liu, and, D. Metaxas, Video Object Segmentation by Hypergraph Cut, CVPR, 2009 [13] D. Lowe, Distinctive Image Features from Scale-Invariant key points, International Journal of Computer Vision, 60(2), pp. 91110, 2004.

[14]

S. J. Sun, D. R. Haynor, and Y. M. Kim, Semiautomatic Video Object Segmentation Using VSnakes, IEEE Trans. on Circuits and Systems for Video Technology. 13(1), pp. 75-82, 2003. [15] N. Vaswani, Y. Rathi, A. Yezzi, and A. Tannenbaum, Deform PF-MT: Particle Filter with Mode Tracker for Tracking Non-Affine Contour Deformations, IEEE Trans. Image Processing, 19(4), pp. 841-857,2009. [16] C. Wang, T. Hassan, N. El-Sheimy, and M. Lavigne, Automatic Road Vector Extraction for Mobile Mapping Systems, XXI Congress, ISPRS, 2008. [17] T. T. Zin, and H. Hama, A Method Using Morphology and Histogram for Object-based Retrieval in Image and Video Databases, International Journal of Computer Science and Network Security, 7(9), pp. 123-129, 2007.