Switching Local and Covariance Matching for Efficient Object Tracking
Junqiu Wang, Yasushi Yagi · InTech eBooks · 2011
matrices only on those pixels that are classified as foreground.Therefore we can track articulated objects.To speed up the global searching process, we use Log-Euclidean metrics Arsigny et al. (2005) instead of the Riemannian invariant metrics Pennec et al. (2006); Porikli et al. (2006) to measure the similarity between covariance matrices.The model update in covariance tracking Porikli et al. (2006) is also expensive.We update the model by computing the geometric mean of covariance matrices based on Log-Euclidean metrics.The computation is simply Euclidean in the logarithmic domain, which reduces the computational costs.The final geometric mean is computed by mapping back to the Riemannian domain with the exponential.Log-Euclidean metrics provide results similar to their Riemannian affine invariant equivalent but takes much less time.We arrange this chapter as follows.After a brief review of previous works in Section 2, we introduce the local tracking method based on foreground likelihood computation in Section 3. In specific, we discuss target representation for local tracking using color and shape texture information in Section 3.1; we describe our feature selection for local tracking in Section 3.2, and our target localization strategy for local tracking in Section 3.3.In Section 4, we apply Log-Euclidean metric in covariance tracking.We introduce a few basic concepts that are important for our covariance matching in Section 4.1.The extended covariance matching method using Log-Euclidean metric is described in Section 4.2.In Section 5, we give the switching criteria for the local and global tracking.Experimental results are given in Section 6. Section 7 concludes the paper. Related workMany tracking algorithms assume that target motion is continuous.Given this assumption, we can apply local tracking algorithms Comaniciu et al. (2003); Isard & Blake (1998); Wang & Yagi (2008b).In the local tracking algorithms, the mean-shift algorithm Comaniciu et al. (2003) aims at searching for a peak position using density gradient estimation, whereas particle filtering techniques Isard & Blake (1998); Rathi et al. (2005); Wang & Yagi (2009); Zhao et al. (2008); Zhou et al. (2006) use a dynamic model to guide the particle propagation within a limited sub-space of target state.Particle filtering tracking algorithms have certain robustness against sudden motions.The mean-shift algorithm can deal with partial occlusions.Tracking can be formulated as template matching Hager & Belhumeur (1998).A target is characterized by a template that can be parametric or non-parametric.The task of a template matching tracking is to find the region that is the most similar to the template.Template matching techniques do not require the continuous motion assumption.Therefore, it is possible to handle occlusions and sudden motions.We will introduce local tracking and global matching techniques.The objective of our algorithm in this chapter it to combine the advantages of the local and global matching techniques. Local trackingThere are many local tracking methods.Tracking was treated as a binary classification problem in previous works.An adaptive discriminative generative model was suggested in Lin et al. (2004) by evaluating the discriminative ability of the object from the foreground using a Fisher Linear Discriminant function.Fisher Linear Discriminant function was also using in Nguyen & Smeulders (2006) to provide good discrimination.Comaniciu et al. Comaniciu et al. (2003) take of the advantage of this method to their mean-shift algorithm, where colors that appear on the object are down weighted by colors that appear in the background.Collins et 120 Object Tracking www.intechopen.comHow to reference In order to correctly reference this scholarly work, feel free to copy and paste the following: Junqiu Wang and Yasushi Yagi (2011).Switching Local and Covariance Matching for Efficient Object Tracking, Object Tracking, Dr. Hanna Goszczynska (Ed.),