Performance evaluation of correlation filters for target tracking Leopoldo N. Gaxiola 1 , Victor H. Diaz-Ramirez 1 , Juan J. Tapia 1 , Pascuala García-Martínez 2 , Andres Cuevas 1 1 Instituto Politécnico Nacional - CITEDI, Ave. del Parque 1310, Mesa de Otay, Tijuana, B.C., 22510, México 2 Departamento de Óptica, Universitat de València, C/Dr. Moliner 50, 46100 Burjassot, Spain ABSTRACT A performance evaluation of several state-of-the-art correlation filters within the context of target tracking is presented. The filters are tested using an introduced algorithm that is adapted online using information of current and past scene frames of the scene. The algorithm achieves a high-rate operation by focussing signal processing on a small fragment of the scene in each frame. The correlation filters are tested using several video test sequences that contain geometric modifications of the target, partial occlusions and clutter. The performance of the tested filters is characterized in terms of detection efficiency, tracking accuracy, and computational complexity using objective metrics. Keywords: Target tracking, correlation filtering, performance evaluation. 1. INTRODUCTION In the last decade, target tracking has received much research interest by the signal processing and computer vision community. Video surveillance, robotics, and human-computer interaction, are examples in where target tracking is required. Target tracking consists in estimation of the trajectory of a target while its moves through a detection zone. 2, 4 The main challenges of target tracking are the presence of high cluttering and additive noise, geometric distortions of the target (rotations and scaling), and nonuniform illumination conditions. Moreover, eventual occlusions of the target must be solved by the tracking algorithm. The use of correlation filters for target tracking has increased in recent years. This is because these filters can estimate with high accuracy the position of a moving target in noisy scenes. Correlation filtering is a template matching approach given by a linear system. In this method, the coordinates of the maximum value in the system output are taken as estimates of the target coordinates within the observed scene. Correlation filters are usually designed by optimization of several performance criteria. These filters can be broadly classified into two main categories; analytical and composite filters. 7 Analytical filters optimize a statistical criterion utilizing mathematical models of the measured signal and the noise. Composite filters are synthesized by combining several training templates, each of them representing a different view of the target that is expected to be present in the input scene. The performance of a composite filter highly depends on the proper selection of the image templates used for training. 10 Recently, Bolme et al, 3 proposed a real-time tracking algorithm based on an adaptive correlation filtering. This proposal yields competitive results with respect to standard tracking algorithms. 5 In Bolme’s approach, a correlation filter (template) is used to detect and locate the target within the scene in each observed frame. The template is updated online (adapted) accordingly with current and past scene observations, and by taking into account intraclass distortions of the target. This algorithm utilizes the Minimum Output Sum of Squared Error (MOSSE) filter. 3 In this work, we present a performance evaluation of several state-of-the-art correlation filters within the context of target tracking. The chosen filters are tested within an introduced adaptive algorithm using several Applications of Digital Image Processing XXXVIII, edited by Andrew G. Tescher, Proc. of SPIE Vol. 9599, 959904 · © 2015 SPIE · CCC code: 0277-786X/15/$18 doi: 10.1117/12.2188433 Proc. of SPIE Vol. 9599 959904-1 Downloaded From: http://proceedings.spiedigitallibrary.org/ on 10/14/2015 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx