1. Golnaz Abdollahian, Cuneyt M. Taskiran, Zygmunt Pizlo, and Edward J. Delp C AMERA M OTION -B ASED A NALYSIS OF U SER G ENERATED V IDEO IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 12, NO. 1, JANUARY 2010

2.  UGV generally has a rich camera motion structure that is generated by the person taking the video and it is typically unedited and unstructured.  The main application of our system is for mobile devices which have become more popular for recording, sharing, downloading and watching UGV  use computationally efficient methods  We propose a new location-based saliency map which uses camera motion information to determine the saliency values of pixels with respect to their spatial location in the frame. Introduction

3.  Global Motion Estimation In the majority of UGV, camera motion is limited to a few operations, e.g. pan, tilt, and zoom; more complex camera movements, such as rotation, rarely occur in UGV our goal here is to be computationally efficient to be able to target devices with low processing power such as mobile devices use a simplified three-parameter global camera motion model in the three major directions Motion-Based Frame Labeling H :horizontal V : vertical R : radial

4.  Template   The iteration stops when a local minimum is found Motion-Based Frame Labeling L1 distance between the 2-D template in the current frame and the previous template

5.  Motion Classification – support vector machine(SVM) is used  We first classify it as having a zoom or not, using the 3-Dmotion vector as the feature vector  SVM classifiers are trained on an eight-dimensional feature vector derived from the parameters H and V over a temporal sliding window. The size of sliding window is different for blurry(N=7) and shaky(N=31)  The frames that are not labeled as zoom, blurry or shaky are identified as stable motion with no zooms. Motion-Based Frame Labeling

7.  Two frames are considered to be correlated if they have overlap with each other  Camera View : a temporal concept defined as a set of consecutive frames that are all correlated with each other.  View boundaries occur when the camera is displaced or there is a change of viewing angle  To detect view boundaries for temporally segmenting the video,we defind the displacement vector between frames i and j as T emporal V ideo S egmentation B ased on T he U se Of C amera V iew

8.  A boundary frame is flagged whenever the magnitude of the displacement vector,, for the current frame and that for the previously detected boundary frame is larger than  There is a constraint that boundary frame can’t be chosen during intervals labeled as blurry segments. T emporal V ideo S egmentation B ased on T he U se Of C amera V iew

9.  A keyframe should be the frame with the highest subjective importance in the segment in order to represent the segment it is extracted from  Since our intention was to avoid the complex tasks of object and action recognition in our system, our keyframe selection strategy was only based on camera motion factor.  The following frames are selected as keyframes :  The frame after a zoom-in  The frame after a large zoom-out  The frame where the camera is at pause  For segments during which the camera has constant motion, all frames are considered to be of relatively same importance. In this case, the frame closest to the middle of the segment and having the least amount of motion is chosen as the keyframe in order to minimize blurriness Keyframe Selection

10. Combine several saliency map to generate the keyframes saliency maps  color contrast saliency  moving objects saliency map  highlighted faces  location-based saliency map Keyframe Saliency Maps and ROI Extraction

11.  Use the RGB color space to generate the contrast-based saliency map  The three-dimensional pixel vectors in RGB space are clustered into a small number of color vectors using generalized Lloyd algorithm (GLA) for vector quantization  Color Contrast Saliency Map Pij and q : RGB pixel value Θ : neighborhood of pixel (i,j) (5*5) d : gaussian distance

12.  To determine the moving object saliency map, we examine the magnitude and phase of macro block relative motion vectors  Relative motion vector for the macro block at location (m,n) :  If relative motion below a threshold values -> assign 0  The motion intensity I and motion phase φ are defined as Moving Object Saliency Map

13.  The phase entropy map, H p, indicates the regions with inconsistent motion which usually belong to the boundary of the moving object  Moving Object Saliency Map the probability of the kth phase whose value is estimated from the histogram

14.  The direction of the camera motion also has a major effect on the regions where a viewer “looks” in the sequence  The global motion parameters were used to generate the location saliency maps for the extracted keyframes Location-Based Saliency Map k H,k V,K r : constant (10,5,0.5) r : distance of a pixel from the center r max : maximum r in the frame

15.  After combining the H and V maps, the peak of the map function is at  The radial map, S R, is either decreasing or increasing as we move from the center to the borders, depending on whether the camera has a zoom-in/no-zoom or zoom-out operation Location-Based Saliency Map

16.  First, the color contrast and moving object saliency maps are superimposed since they represent two independent factors in attracting visual attention  Faces are detected and highlighted after combining the low level saliency maps  The location-based saliency map is then multiplied pixel-wise with this map to yield the combined saliency map Combined Saliency Map

17.  A region growing algorithm proposed is used to extract ROI from the saliency map  Fuzzy partitioning is employed to classify the pixels into R 1 : ROI and R 0 : insignificant regions  seed selection  1) the seeds must have maximum local contrast  2) the seeds should belong to the attended areas Identification of ROIs

18. Experimental Results

20. left Zoom-out Zoom-in

21. Experimental Results

22.  UGVs contain rich content-based camera motion structure that can be an indicator of “importance” in the scene  Since camera motion in UGV may have both intentional and unintentional behaviors, we used motion classification as a preprocessing step  A temporal segmentation algorithm was proposed based on the concept of camera views which relates each subshot to a different view  We use a simple keyframe selection strategy based on camera motion patterns to represent each view  we employed camera motion in addition to several other factors to generate saliency maps for keyframes and identify ROIs based on visual attention Conclusion