Three-dimensional Object Tracking in RGB Datasets

There are many works with many progresses using RGB-D on object tracking when long-term occlusion occurs. However, object tracking needs a higher requirement on hardware, like RGB-D cameras. To solve this problem, this paper proposes a novel depth information fusion tracker (DIFT) which handles occlusion by the following series of steps on general cameras. First, object occlusion is recognized during object detection. Second, a dynamic tracking model is established and updated according to whether the object is occluded. Third, the tracking model is refined by eliminating occluded regions. An extensive quantitative evaluation on public video sequences shows that the proposed method is robust and outperforms widely used trackers such as Kernelized Correlation Filter and Tracking-Learning-Detection.