Robust Visual Tracking Using Randomized Forest and Online Appearance Model

Abstract

We propose a robust tracker based on tracking, learning and detection to follow an object in a long term. Our tracker consists of three different parts: a short term tracker, a detector, and an online object model. For the short-term tracker, we employ the Lucas Kanade tracker to keep following the object frame by frame. Meanwhile, the sequential randomized forest using a 5bit Haar-like Binary Pattern feature plays as a detector to detect all possible object candidates in the current frame. The online template-based object model consisting of positive and negative image patches decides which the best target is. Our method is consistent against challenges such as viewpoint changes, various lighting conditions, and cluttered background. Moreover, our method is efficiently able to reacquire the object efficiently even after it’s out of view or in total occlusion. We also propose an efficient way to extend our tracker for multiple faces tracking application. Extensive experiments are provided to show the robust of our tracker. Comparisons with other state-of-the-art trackers are also demonstrated.