Single Object Visual Tracking

In this project, we are working on visual tracking which is an important field of computer vision. More specifically, visual tracking is to automatically follow a specific target determined in the first frame of a video under challenging scenarios such as fast target motion, occlusion and deformation. Recently, machine learning methods have been successfully employed in tracking algorithms to improve their performance. Thus, we have focused on different methods of machine learning in the context of visual tracking.

After successfully proposing two novel visual trackers named “DCPF” and “DCPF2”, we are completing the third one. In the first tracker, we combined a particle filter with a CNN and a correlation filter. Generally, CNN-correlation trackers use the convolutional features to generate the target model in the first frame and updating it in the next frames. Making comparison between the target model and the convolutional features extracted in each frame results finding the target position. Particle filters improve the tracker’s performance in challenging scenarios. In DCPF2, we extended the particle filter to estimate the bounding box.  Lastly, we developed a adaptive correlation filter in which different number of the target models are generated in each frame, based on all high-likelihood particles