Adaptive Face Tracking Based on Online Learning

Abstract

Object tracking can be used to localize objects in scenes, and also can be used for locating changes in the object’s appearance or shape over time. Most of the available object tracking methods tend to perform satisfactorily in controlled environments but tend to fail when the objects appearance or shape changes, or even when the illumination changes (e.g., when tracking non-rigid objects such as a human face). Also, in many available tracking methods, the tracking error tends to increase indefinitely when the target is missed. Therefore, tracking the target objects in long and uninterrupted video sequences tends to be quite challenging for these methods. This work proposes a face tracking algorithm that contains two operating modes. Both the operating modes are based on feature learning techniques that utilize the useful data accumulated during the face tracking and implements an incremental learning framework. To accumulate the training data, the quality of the test sample is checked before its utilization in the incremental and online training scheme. Furthermore, a novel error prediction scheme is proposed that is capable of estimating the tracking error during the execution of the tracking algorithm. Furthermore, an improvement in the Constrained Local Model (CLM), called weighted-CLM (W-CLM) is proposed that utilize the raining data to assign weights to the landmarks based on their consistency. These weights are used in the CLM search process to improve CLM search optimization process. The experimental results show that the proposed tracking method (both variants) perform better than the comparative state of the art methods in terms of Root Mean Squared Error (RMSE) and Center Location Error (CLE). In order to prove the efficiency of the proposed techniques, an application in yawning detection is presented. 1