General Single Object Tracking Algorithm Under Complex Dynamic Changes

General single object tracking is an active research topic in the field of computer vision,and has also received extensive attention in practical applications.Different from the object detection task based on static pictures,video-based object tracking task needs to deal with complex and dynamic changes in targets.In addition,there is no specific target category for tracking task,so offline network training and initial modeling of the target are very difficult.The above two points make it difficult for the tracker to maintain continuous tracking of the target.This thesis will proceed from the above two difficulties to study the problems in the three mainstream general single object tracking algorithms under complex dynamic changes,and propose corresponding improvements to improve the performance of the algorithm.The first research is the single object tracking algorithm based on correlation filters.Correlation filters algorithms generally have the problem of boundary effects.In order to suppress the boundary effects,a cosine window is usually added to the search area.However,the cosine window will limit the search field of view.When the target has a large displacement in space due to complex dynamic changes,the tracker will lose the target.In order to solve this problem,this paper proposes a single object tracking algorithm based on correlation filters with spatial attention module.By modeling the target and the background,the tracker is provided with a priori spatial information of the target during the inference phase,eliminating the problem of the cosine window restricting the tracker's field of view.At the same time,an adaptive multi-feature fusion strategy is proposed to enhance the expression ability of traditional features.Secondly,we study single object tracking algorithm based on object detection.Object detection,as a basic field of computer vision,has produced many excellent algorithms on its tasks.Attracted by their powerful classification capabilities and accurate regression capabilities,more and more single target tracking algorithms learn and draw on object detection techniques.In order to fully exploit the potential of target detection in the field of target tracking,two problems need to be solved: the difference between the two tasks' definition of the target and the difference in the time dimension.In order to eliminate the above two differences,we propose a "Detection to Tracking" framework.This framework builds a general-to-special feature extraction network to solve the classless prior problem for single object tracking tasks.At the same time,we proposed a long-term and short-term timing strategy to adapt to the target's appearance change.As far as we know,this framework is the first to study the transplantation of target detection algorithms to single target tracking tasks.It provides a novel solution to the single object tracking problem.Finally,it is the Siamese network based trackers.The Siamese network is a deep network structure which is widely used in the field of single object tracking with the development of deep learning.The Siamese network uses the idea of template matching to offline training the convolutional neural network.Although the offline training phase provides a large number of training samples,we find that they are full of redundant and inefficient samples.These samples make it difficult for the Siamese network to adapt to the target's appearance variations and background distractions.We propose a data augmentation method based on intrinsic responses to improve the training process of Siamese networks.We use the output of the network to derivative the network input to obtain an intrinsic response map that reflects the influence of the internal characteristics of the network on the network result,and then use the endogenous response map to synthesize intrinsic samples based on target appearance variations and background distractions.A large number of experiments prove that intrinsic samples can effectively improve the performance of Siamese networks.