Visual Tracking with Motion Regularizer and Adaptive Update

Abstract

Although the discriminative method has shown strong performance in the aforementioned works, there are still some problems that can’t be solved properly, such as the template’s scale problem and the target’s feature representation problem. While Convolution Neural Network has shown good performance in recent image related problems. the two method lack better deep integration. For a CNN that is trained and tuned on a large-scale image repository, our framework uses the output of the middle layer as the feature representation, because they show good performance in many visual related tasks. We noted that it is of great significance to obtain good and effective proposal image blocks about the target location to improve the practicability of the tracker. We adapt location proposal network to our tracking task. In addition, support vector machine is used to classify features. A new model updating strategy is proposed. Our scheme shows good performance on large dataset.