A Robust Framework for Accurate Face Tracking Ali Ahmed * IT Department, Faculty of Computers and Information, Menoufia University, Shebin El-kom, Egypt. * Corresponding author. Tel.: +966541924606; email: ali.ahmed@ci.menofia.edu.eg Manuscript submitted March 3, 2017; accepted May 17, 2017. doi: 10.17706/jsw.12.5.326-338 Abstract: Face detection and tracking are of the most challenging problems of the object tracking field because of the large variability of faces and facial expressions. In this paper, two different algorithms for face tracking based on unscented Kalman filter (UKF) are proposed. The first proposed algorithm is UKF based on Viola-Jones algorithm. Viola-Jones is extremely fast feature computation, efficient feature selection, and scale and location invariant detector. The second proposed algorithm is UKF based on mean shift using the corrected background weighted histogram (CBWH) scheme. This scheme can effectively reduce background's interference in target localization and consequently can guarantee accurate localization of the target. The tracking step is completed using UKF that can estimate the next state with a high level of accuracy. So the two proposed algorithms are used to enhance the solution of face tracking problems. The performance of the two different proposed algorithms is evaluated with other well-known face tracking algorithms. Key words: Mean shift, corrected background weighted histogram, unscented kalman filter, Camshift. 1. Introduction The human face is important to our identity. It plays a major role in everyday interaction, communication, and other routine activities. Recently, Face detection step is mainly concentrated on searching for faces on a given image. If a face exists, immediately returns the image location and content of each face. The goal of this approach is to emerge the best face detection approach that satisfies the need for real-time hardware implementation. The face detection problem is also a very challenging where it needs to account for all possible appearance variation due to a change in illumination, facial features, and partial or full occlusions. Also, the challenging is when needed to detect faces that appear at a different scale, or out of a plane rotation, or with in-plane rotation. Although, all these difficulties, tremendous progress has been made in the last decade, and many systems have shown impressive real-time performance. As this problem is the first step of any face processing system, it is used in many applications of face recognition and tracking, facial feature extraction, attentive user interfaces, gender classification, clustering, digital cosmetics, biometric systems, and Human Computer Interaction system, demographic classification, and surveillance system. Kalman filter is a well known and widely used optimal estimator for linear systems. This filter predicts the location of a moving object based on its previous state. From a series of noisy measurements, Kalman filter is a recursive adaptive filter that estimates the state of a dynamic system. It has been used for an extensive range of applications in areas such as signal, image processing, and control design. Unfortunately, Kalman filter is widely used to reduce the dimensionality error only when the systems in the real world are linear. A common solution to cope with major problem is to linearize the system first before applying the filter, resulting in a new approach called an extended Kalman filter (EKF). This linearization when does, however, pose some other problems, for 326 Volume 12, Number 5, May 2017 Journal of Software