서강대학교

초록/요약

Computer vision has experienced many advances over the recent years. As such, it started to rapidly find its way into many commercial and industrial applications, with disciplines like visual object detection and tracking being very prominent. Visual object tracking addresses the problem of following an object in a video scene, and it is still a very challenging task, with many datasets and benchmarks that are not still saturated and present difficulty even for newly developed methods. Moreover, new technologies such as Unmanned Aerial Vehicles (UAV) presents new challenging scenarios that were not consider before, and helps the discipline to continue improving. Recently, with the help of challenges of the likes of the Visual Object Tracking (VOT) Challenge, huge efforts have been put in study the case of generic object trackers, were there is no prior information about the object we want to follow throughout the scene. Research done has led to the development of tracking algorithms based on the hot topic of deep learning, specifically Convolutional Neural Networks (CNN); and trackers based on learning a filter from the object appearance using correlation operation, referred as Correlation Filter Trackers. This thesis studies the problem of single object tracking (SOT) in its generic form. This works presents a combination tracking algorithm of an state-of-the-art CNN-based tracker called MDNet, with correlation filter based trackers named fDSST, DPCF and ECO, with the goal of improving the performance and/or speed of MDNet, to help push the limits of state-of-the-art algorithms. The CF-based algorithms are used as “guides” to help MDNet choose a proper set of candidate boxes to evaluate, in order to determine the new location of the target. Experiments shows that with the help of correlation filter based trackers the accuracy performance of MDNet can be improved. However, as a result of adding another component to the tracking framework, the processing time is increased. Based on these experiments, we are able to identify shortcomings of current MDNet framework, providing possible ways to improve it and directions for future works.