Improvement of Human Tracking Based on an Accurate Estimation of Feet or Head Position

In this paper a method is presented to estimate the position of feet/head of objects in various camera views. In this method, first, all objects in the scene are detected using the background subtraction. Then, human and non-human objects are separated via the support vector machine (SVM) that is trained based on local binary patterns (LBP) features. The basic idea of the next step of this work is that the feet/head of an object are the group of pixels that are projected to small region on ground/top plane by corresponding homography matrix. This idea is expressed via an optimization problem which avoids partitioning out small group of pixels. Experimental results show that the proposed methods can improve the accuracy of the object tracking.