| The successful launch and operation of video satellites has expanded the frontiers of remote sensing ground observation technology.With the ability of video satellites to "gaze" to observe the ground area and record video information,new research fields and application directions have emerged.Object tracking technology is an important research task in computer vision.It can continuously track the moving state of the target of interest in the video.The combination of object tracking technology with satellite video data can provide essential help for applications such as traffic flow monitoring analysis and emergency response.Compared with general video,satellite video has lower resolution,more complex background,and objects in it have smaller size.These characteristics make it less distinguishable between different targets or target and background,thus bring certain challenges to object tracking tasks.After long-term development of object tracking technology,many effective technical solutions have been formed.Among them,correlation filter algorithms have become mainstream due to their accuracy and high speed.With the successful application of deep learning methods in the field of computer vision,object tracking algorithms based on deep neural networks are also becoming more mature and fruitful.According to the knowledge in object tracking domain,and considering the characteristics of satellite video,this article has conducted in-depth research on designing an object tracking algorithm that can be effectively applied to satellite video.Based on the basic architecture of the convolutional neural network,this article proposes a tracking algorithm that uses a convolutional regression network model as the core component and solves model parameters through gradient descent technique.By training with enough real samples,this model could improve the deficiency of correlation filter algorithms,and increasing the tracker's discriminative power in satellite video scenes.In feature extraction,a pre-trained convolutional neural network model is used instead of traditional hand-crafted low-level features to obtain high-level target representations.Further,this article proposes to extract appearance features and motion features separately,and combine them through adaptive weights to make reasonable use of the complementary information between the two type of features.This strategy could overcome the shortcomings of using only a single feature,improve the robustness,and effectively solve the tracking drift problem caused by similar targets nearby.This article uses satellite video data to establish a complete test data set according to the common evaluation benchmarks for object tracking,and experimentally validates the proposed algorithm.Comparing the evaluation results of common object tracking algorithms with our algorithm,it shows that this algorithm has the superior performance of tracking in satellite video scenes.Through a large number of ablation experiments and visualization analysis of the tracking results,it is proved that the design of the algorithm is reasonable. |
