Visual Target Tracking Based On Optimized Ensemble Learning And Spatial Correlation Filter

There are many applications that use the camera to track objects in the environment,such as intelligent monitoring,healthcare,intelligent transportation,and human-computer interaction.Visual target tracking is a very active research area and is also a challenging task.Nowadays,one of the most popular approaches to visual target tracking is tracking-by-detection.Tracking-by-detection relies on a discriminant learning model to learn a binary classifier online,which separates the target from the background.The goal of this method is to design a robust classifier to dynamically capture the differences between the target and the background,adapting the variations in illumination,viewpoint,and pose.In this thesis,the commonly used tracking algorithms are studied in depth,some difficult problems are analyzed and some solutions are put forward.The main achievements of the thesis are as follows:(1)An ensemble learning tracking method based on logistic regression model is proposed.The algorithm predicts the target with simple and fast weak classifiers.In order to overcome the performance defects of the weak classifier,it is selected and integrated by logistic regression.The algorithm improves the accuracy of tracking on the basis of simple features(Haar-like features)and rough classifiers.(2)A long-term tracking method based on correlation filtering with an adaptive target response is proposed.In order to avoid tracker drift,the background information near the target is used as a priori to construct a translation filter with an adaptive target response.A 1-dimensional scale filter can be applied at an image location to estimate the target size.At the same time,a robust SVM detector is used to handle tracking exceptions.Through the collaboration of the above three modules,the algorithm can deal with complex scenes in tracking.Experiments show that the performance of the algorithm is better than other similar algorithms.(3)A structured output tracking method based on correlation filter is proposed.In order to solve the boundary effect of the correlation filter,the samples are trimmed in the spatial domain,the deformed parts in the sample are removed,and the samples are defined in a structured way to better describe the information contained in the sample.The method for the sample processing allows the filter to learn the deterministic features of the target and the background.Experiments show that the algorithm can adapt to the change of target and background,and has good performance.