CNN feature 기반 correlation filter을 사용한 물체 추적

초록

Object tracking is a challenging work in computer vision as appearance model changes caused by occlusion, background clutter, fast motion, illumination variation and deformation. In this paper, CNN-based features extracted from different convolutional layers are used to express appearance model to improve tracking accuracy. The deepest convolutional layer of CNN keeps semantics information of object, which is robust to significant appearance variations, combining with early layers which retaining location information to extract features. Then correlation filters are trained in different convolutional layers to encode object appearance and get confidence scores. The highest one is chosen as predicted target positions on each layer. Lastly, final target position is fixed using non-maximum suppression. Our algorithm is evaluated by benchmark 2013, the distance precision (DP) rate at threshold of 20 pixels can arrive 79.8% and overlap success (OS) rate at an overlap threshold of 0.5 is 54.3%. Experimental results demonstrate that the performance of our method is much better than state-of-the-art methods in terms of accuracy and robustness.