Adaptive visual tracking with failure detection

Most traditional visual tracking algorithms ignore the importance of tracking failure detection which is helpful for occlusion handling. To address this problem, a robust visual tracking algorithm combining the tracking model and the failure detection strategy is proposed for the needs of practical applications. In order to tracking the position and scale of the target with high speed and accuracy, a method that estimates the target scale after localization is proposed. When tracking, online SVM is used to estimate the location of the target. Then, variable patches with different sizes are sampled at the estimated location to get scores from the SVM, among which the one scored highest is selected to be the target object. Moreover, a failure detection strategy is introduced based on Log-polar transformation, Gaussian mixture model, and time-domain filtering. For this failure detection model, Log-polar transformation is adopted as it is much more sensitive to drift than scaling and rotation. Thus, whether the target has disappeared or not can be easily judged by the degree of drift using Log-polar transformation and Gaussian mixture model. The time-domain filtering scheme is used to reduce the false-alarm probability, and makes the failure detection more robust. Experimental evaluation on standard visual tracking datasets demonstrates the superiority of our algorithm by comparing with state-of-the-art trackers such as Struck and TLD.