REALTIME OBJECT-OF-INTEREST TRACKING BY LEARNING COMPOSITE PATCH-BASED TEMPLATES

1. 2010 International Conference on Image Processing REALTIME OBJECT-OF-INTEREST TRACKING BY LEARNING COMPOSITE PATCH-BASED TEMPLATES Yuanlu Xu, Hongfei Zhou, Qing Wang*, Liang Lin Sun Yat-sen University, Guangzhou, China INTRODUCTION • CPT model is initialized by selecting image patches from the tracking window, which maximizes the difference with the background. • Tracking Algorithm • To infer the tracking target location, as shown in Fig. 4, we independently find the best match of each template (green rectangle) within the surrounding area (blue rectangle). And compute the location with the following formulation: • To maintain the CPT model online, we propose a new maintenance algorithm by picking new patches and fusing them into CPT model, as shown in Fig. 5. Fig. 6. Sampled results of our tracking methods, target with severe body variations and large-scale occlusions. • Objective • To tracking people with partial occlusions or significant appearance variations, as Fig. 1 illustrates. • Contributions • A novel model maintenance approach for patch-based tracking. • A more concise and effective texture descriptor. Girl Fig. 3. Illustration of CS-LBP descriptor, CS-LBP operator compares the intensities in center-symmetric direction. Face Occlusion Fig. 1. Difficulties of tracking: partial occlusions (left two), significant appearance variations (right two). Camera 1 OUR METHODS Camera 2 Fig. 4. Finding the best match of each template. • Composite Patch-based Templates (CPT) Model • We extract image patches with size from the tracking window and the up, down, left and right 4 sub-regions around the tracking window, forming the patch set of the tracking target and the patch set of the background • , respectively. • For each image patch, different types of features are applied to capture the local statistics. • Histogram of gradient (HOG), to capture edges. • Center-symmetric local binary patterns (CS-LBP), as illustrated in Fig. 3, to capture texture. • Color histogram, to capture flatness. where indicates the offset of each template, the discriminability weight. Fig. 7. Quantitative comparisons Constructing the candidate template set with a new frame and the inferred target location, we obtain two feature sets , from the inferred tracking window and the background, respectively. The new CPT model Fusing in an excess model CONCLUSION Candidate template set Re-selecting • The novel maintenance approach selects effective composite templates from the fusion of the matching templates and the candidate set, which outperforms other state-of-art algorithms in tracking targets with various challenges. • CS-LBP descriptor is an effective dimension-reduced texture descriptor. By thresholding the matching distance, we can obtain the matching templates in two successive frames. Fig. 2. Illustration of CPT model. REFERENCES Matching template set Fig. 5. Find the most discrimination patches in matching templates and candidate templates to maintain the CPT model online. • [1] X. Liu et al, “Representing and recognizing objects with massive local image patches,” Pattern Recognition, vol. 45(1), pp. 231–240, 2012. • [2] Y. Xie, L. Lin, and Y. Jia, “Tracking objects with adaptive feature patches for PTZ camera visual surveillance,” ICPR, pp. 1739–1742, 2010. • [3] X. Liu, L. Lin, S. Yan, H. Jin, and W. Jiang, “Adaptive object tracking by learning hybrid template on-line,” TCSVT, vol. 21(11), pp. 1588–1599, 2011. EXPERIMENTS • We collect 4 test videos to verify our approach, two video of the human face from Multiple Instance Learning (MIL) and two surveillance videos from the internet. A number of sampled tracking results are shown in Fig. 6. • We compute the average tracking error with manually labeled groundtruth of the tracking target location and compare with two state-of-the-art algorithms: MIL and Ensemble Tracking, as shown in Fig. 7. FOR FURTHER INFORMATION • Please contact merayxu@gmail.com, linliang@ieee.org.