Localization-based Visual Tracking with Convolutional Neural Networks

This paper presents a novel framework for the visual tracking problem. This framework predicts the exact location of the object using a regression. In this work, we first select an approximate region based on object location in the previous frame and then predict the exact location of the object in the current frame by a deep convolutional network that its last layer replaced with a regression. The entire network gets updated due to the occurrence of various challenges during the video. We evaluate our work using 8 challenging benchmark video sequences and shows a significant improvement over state-of-the-art approaches.