A Submodular Optimization Tracking Algorithm Based On Discriminative Feature Regression

With the great progress in computer technology in recent years, image and video processing, computer vision has been an attracting research field. Among visual problems, the tracking of moving objects in video has a wide application in lots of fields, such as human-computer interaction, video surveillance, video communication, traffic control, social security and industrial production. There has been many tracking methods proposed by researchers, which achieve favorable results in tracking. However, due to the variation of pose, scale in practical situation, shape distortion, illumination or occlusion, the appearance changes a lot in tracking process, so it's still a challenging task to track the object with robustness. The key point is to construct a reliable and efficient appearance model, which could catch the appearance changes and describe the target under these variations, so we could track the target with robustness.In earlier research, we found there is some dependency relationship between the feature vector and target confidence in subarea, we could catch the appearance feature under disturb factors according to that. So we propose a submodular tracking method based on discriminative feature regression, this method is from the perspective of mid-level vision, we construct a data set which consists of superpixels feature vectors and corresponding confidence, then train an appearance model by regression. In order to improve the accuracy of the regression model, we introduce the submodular function to select more representative cues of foreground, and optimize the preliminary prediction confidence map of regression model. Finally, we take the background information into consideration, under the Bayesian platform, the tracking task is formulated as obtain the best candidate by maximizing the posterior probability estimation on confidence map. We test our tracking method on several public datasets, comparing to the state-of-art tracking method, our method is efficient and robust under appearance changes.