Online learning of cascaded classifier designed for multi-object tracking

Visual multi-object tracking is an important task within the field of computer vision. The goal of this paper is to track a variable number of unknown objects in complex scenes automatically using a moving and un-calibrated camera and it devotes to overcome the challenging problems including illumination and scale variations, viewpoint variations and significant occlusions, etc. In this paper, a binary representation containing color and gradient information is utilized to obtain unique features so that the objects can be easily distinguished from each other in the feature space. In addition, an online learning framework based on a cascaded classifier which is trained and updated in each frame to distinguish the object from the background is proposed for long-term tracking. The experimental results on both quantitative evaluations and multi-object tracking show that this approach yields an accurate and robust tracking performance in a large variety of complex scenarios.