Extended Deep Regression Network Based Visual Object Tracking

Computer vison has many sub-branches,and visual object tracking is one of the most important branches.Tracking task is a kind of multi-domain project which involves many kinds of knowledge,also connect with diversified interference.Deep learning knowledge is often used in computer vision because of their ability to learn from large number of training data to improve performance.However,most visual object trackers are trained from scratch online and do not take the advantages of the large amounts of videos and images that are available for offline training.Tracking algorithm are supposed to track object in real-time while using to solve problems,so that we raise the Extended Deep Regression Network based Visual Object Tracking(EDR)to develop a tracker.Our tracker does not use a complex feed-back network with no online training required or multi-layer network architecture to ensure the accuracy.We train a neural network for tracking in an entirely offline manner.The tracker learns a generic relationship between object motion and appearance.And the tracker can be used to track novel objects that do not appear in the training set because our training procedure can take the advantage of not only a collection of videos but also a set of still images that are each labeled with the location of an object.At the same time the EDR try to take the advantage of salience-map of the search window to improve the performance while the face the occlusion and complex background.We get the salience map of the search window in the current frame,then we could get the lightest point in the map,which would be the referenced point to rectify the predicted object region.We test our network on a standard tracking benchmark to demonstrate our tracker's state-of-the-art performance.Further,our performance improves as we add more videos to our offline training set.