Object Tracking Algorithm Based On Anchor-free Correlation Filtering

Visual object tracking has always been one of the most basic tasks in the field of computer vision,and it has a wide range of application scenarios in smart cities,smart security,human-computer interaction and other fields.With the introduction of deep learning technology,the performance of object tracking algorithms has been greatly improved.The current mainstream target tracking algorithms are all improved on the basis of the twin network.In the siamese network,accurate and fast tracking can be achieved.But in some complex scenarios,there are still some challenges which need to be improved.For example,in the scene where the foreground and background are similar,the tracker will not distinguish the foreground from the surrounding background well;when the target is occluded,the tracker will lose track of the target;when the target does not appear in the training set,The target will not be detected by the device.In addition,most of the current mainstream target tracking algorithms use an anchor frame-based tracking mechanism.Although the introduction of the anchor frame mechanism will increase the accuracy of target position detection,it also brings too many parameters and requires multiple training iterations and optimizations.In order to obtain an excellent result.Therefore,single target tracking is a very challenging task.The main work and innovations of this paper are as follows.First,in order to solve the occlusion problem in the object tracking,this paper proposes a stream-based tracking algorithm.The algorithm can detect occluded objects based on the interrelationship between frames,so as to solve the occlusion problem in the tracking process.The channel attention mechanism is introduced in the feature extraction network to supplement the semantic information in the feature.Second,this paper designs a new high-resolution network that can keep the input resolution and output resolution the same.In the network design stage,this paper analyzes the reason why the network depth cannot be too deep in the correlation filter network.Experiments show that the target tracking algorithm based on correlation filtering must ensure that the resolution of input features and output features are as consistent as possible.Third,in the dataset processing stage,the dataset used in the traditional siamese network is analyzed.This paper proves that the imbalance between the non-semantic background and the semantic background in the trainset is the main obstacle in the learning process.This paper introduces the large-scale dataset ImageNet and the target detection dataset COCO to expand the proportion of positive sample pairs.Fourth,in order to make the obtained features more robust,this paper proposes a method of fusing high-level features with low-level features using feature fusion methods and calculating their similarity.In traditional target tracking networks,most of them only use high-level features or low-level features.The low-level features have higher resolution and contain more position and detailed information,but their semantics are lower and noise is more.High-level features have stronger semantic information,but the resolution is very low,and the perception of details is poor.Finally,this paper conducts a large number of experiments on the proposed algorithm with OTB-2015,VOT2018,UAV123 and LaSOT benchmark datasets.Multiple experiments show that the proposed algorithm has better performance than other methods when dealing with challenging scenarios.During the test,the tracker can run at a speed of more than 56 FPS without extra skills,which satisfies real-time performance.