MACHINE PERCEPTION AND ROBOTICS GROUP

機械知覚&ロボティクスグループ
中部大学

We will present in this paper a method based on invariant feature points with two attributes for robust object tracking against various appearance changes and occlusion. The first attribute, referred to as the affiliation attribute, determines whether each feature point belongs to the tracked object target region or to the background region. The use of this affiliation attribute on each feature point helped us eliminate the feature points in the target region that are affiliated to the background region. It also helped us determine whether occlusion had occurred or not. The second attribute, referred to as the memory attribute, determines the memory term of a feature point through its appearing frequency. The use of this memory attribute helped us minimize the model degradation of long-term stored feature points due to various appearance changes. Through the experiments we performed, we will show that these two attributes can help prevent tracking failures during appearance changes and occlusion where conventional methods are unsuccessful.