Integral Image Based Probabilistic non-linear Subspace Visual Object Tracking

Abstract

Visual tracking is the process of estimating the location of an object in consecutive video frames. It is a major problem in the domain of computer vision with large set of applications such as tra c monitoring, robot planning, surveillance, vehicle navigation and human computer interaction. The problems related to object tracking are complex object shapes, noise in the image, occlusion, and background clutter. The aim of this work is to develop such technique which can track object in a complex environment. This work presents a visual object tracking algorithm using an eigenspace representation. Object appearance and spatial information is learned from a single template using a non-linear subspace projection to arrive at a eigenspace representation. This non-linear subspace representation provides a robust and compact representation of the object. Localization is performed using a similarity measure in non-linear eigenspace representation. A probabilistic search strategy, based on particle lter, is employed to nd the region of an object in each frame of the video sequence that best models the target object in the subspace representation. Particle lter estimates the posterior distribution using weighted samples. Increasing the number of samples increases the estimation accuracy at the cost of increased computations. We, therefore v vi propose a novel kernel subspace integral image framework, which allows the tracker to densely sample the state space without loosing computational ef- ciency. The proposed tracker is tested on number of challenging sequences to demonstrate the performance.