Robust Visual Tracking With Deep Convolutional Neural Network Based Object Proposals on PETS

Gao Zhu, Fatih Porikli, Hongdong Li; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2016, pp. 26-33

Object tracking has been widely used yet still a challenge for surveillance as the drastic size change, deformation and occlusion present. While it is hard to design such an online classifier that adapts to all those changes, in this paper, we employ an object proposal network to generate a small set of bounding box candidates. In a new frame, only these "object-like" candidates are necessary for the classifier to test, which excludes spurious false positives. We also use them to update and improve the discriminative power of the classifier as those proposals are likely to be the background distractions. The novelly proposed approach is robust to object deformation and size change as they are handled naturally during the object proposal stage. We evaluate it on the PETS 2016 dataset, comparing to state-of-the-art trackers.