Object Tracking Research Based On Attention Mechanism

Visual object tracking,which has been a fundamental research topic of computer vision for decades,is studied by researchers from all over the world and has many important applications,such as automatic driving,intelligent traffic monitoring,UAV and so on.Given the initial state of the target in the video,the task of object tracking is to estimate the position and size of the target in the following video frames.In the process of tracking,the appearance of the target is significantly different from the initial state due to scale change,deformation,occlusion and other factors.It is the critical problem that achieving accurate tracking with the influence of these factors.With the development of deep learning and attention mechanism in computer vision,deep learning based trackers show great potential.However,the attention models in current tracking methods do not fully consider those factors,which results in limited improvement.In this thesis,we carefully analyze the difficult factors above and build an accurate model of the target with the attention mechanism and the deep learning based tracking architecture to improve the tracking performance.The main works of this thesis are as follows:(1)We propose a novel dual self-attention based Siamese tracking method.In this method,we propose a dual attention module,which enhances the representation of the target from the spatial and channel aspects,to address difficult factors,e.g.,deformation.In this module,we adopt the adaptive weights and residual connection to select attention features.We combine our module with Siamese tracking network to implement our tracker and conduct many experiments.The experimental results show that our tracker achieves great tracking performance.(2)We propose a novel multi-scale selective attention based tracking method.In this method,we propose a multi-scale selective attention module,which models the target from both small and large scales,to address difficult factors,e.g.,scale change.In the module,we adopt a mixed feature selection strategy to fully integrate the effective parts of different scale features.We implement our tracker based on segmentation tracking network with our module and conduct many experiments.The experimental results show that our module significantly improves the tracking performance and our tracker outperforms many state-of-the-art trackers.