Least squares estimation-based adaptive observation model for aerial visual tracking applications

Abstract

In aerially captured video, low resolution, very small object size and poor image quality may pose tracking as very difficult problem. Based on fundamental principle that single object class lie on a linear subspace, we propose a least squares estimation-based linear model which represents probe candidate as class specific galleries. In this method, tracking is considered as a binary classification problem, which consists of target and background class. Inclusion of background templates improves the discriminative strength of the object model. By projecting target patches on estimated target coefficient vector, all candidates are reconstructed. The squared sum of reconstruction error is used as an observation likelihood function to determine the maximum a posteriori MAP estimation for current frame. To contain the target appearance change during tracking online multiple dictionary learning strategy is proposed. Experiment results of proposed method shows better tracking accuracy and computation when compared to other representative tracking methods.