Anchor-Free Tracker Based On Space-Time Memory Network

Visual object tracking is one of the basic open issues in the field of computer vision,and its task is to predict the position and status of the tracked object online based on its appearance characteristics.Due to various interference factors in the real world,it becomes complex and difficult for the tracker to learn accurate appearance models of tracking objects.For example,the objects may encounter problems such as occlusion,appearance deformation,and interference from similar objects during movement.For these difficult challenges,most existing trackers cannot provide good solutions.The main research contribution of this paper is to propose an anchor-free tracker based on space-time memory network(ATSMN).The specific research contents are as follows.(1)A novel object tracking framework based on space-time memory network is proposed to solve the difficult problems of appearance deformation,occlusion,and similar object interference in object tracking tasks.This tracking framework combines space-time memory network and Transformer network structure for the first time.The overall network framework is not only concise and efficient,but also has strong adaptive ability.Experiments on several large public datasets verify the progressiveness and effectiveness of this algorithm,and meet the real-time tracking requirements.(2)A memory frame feature fusion network is proposed to sovle the problem that traditional trackers based on space-time memory network do not really utilize temporal information.By sequentially extracting multiple memory frames stored in the space-time memory network and adaptively weighting and fusing their features,the network can further extract the temporal context information contained in the memory frames on the basis of obtaining rich appearance information of the tracking object,and obtain better memory frame fusion features.Experimental results show that the proposed memory frame feature fusion network can better guide the tracker to locate objects by utilizing the temporal information in the memory frames.(3)A concise feature cross fusion network based on Transformer is proposed to sovle the problem that feature fusion methods for related operations in traditional siamese network can lead to loss of semantic information and lack of global information.The feature fusion network composed of a simplified Transformer structure can effectively combine the memory frame features with search frame features,enabling the tracker to obtain better classification and regression results while reducing the amounts of parameters.Experimental results show that the proposed Transformer feature cross fusion network is reliable and effective,and can significantly improve the performance of the tracker.Through the coordination and integration of the above innovative efforts,the object tracking algorithm proposed in this paper can achieve accurate,robust,and real-time object tracking.Extensive testing experiments on multiple challenging benchmark datasets have shown that ATSMN achieves SOTA level performance.