| Object tracking is one of the hot issues in the field of computer vision,and is used in various fields of life widely.With the development of science and technology,how to quickly retrieve and track massive video data has become a hot and difficult point in the research of artificial intelligence development.Researchers and scholars have proposed a series of solutions from different perspectives on the problems of existing target tracking algorithms.The traditional video target tracking algorithm is based on manual selection of features.The tracking effect depends largely on the prior experience of the algorithm developer.The design of the algorithm is subjective,the robustness of the algorithm is poor,and it can only be used in specific scenarios.While deep learning technology with its powerful feature automatic extraction capability,the robustness and generalization of the target tracking algorithm is greatly improved.However,due to the complexity of the background and the changes of the object itself,object tracking is still a very challenging research hot spot.The main work of this thesis is:(1)A tracking algorithm for feature fusion is proposed to enhance the anti-interference of the algorithm and improve the discriminative ability of the algorithm.In this paper,the application of appearance features in target tracking algorithm is studied.The color features in the global features and the SURF features in the local features are extracted.Combined with the deep features of the full convolution twinning network,some of the wrong candidate region frames are eliminated,and the discriminating ability of the algorithm is improved effectively.(2)Since the algorithm uses offline tracking,the tracking target is easily lost.Aiming at the shortcoming,a long-term tracking model is proposed.When the fully-convolution siamese networks loses the target tracking failure,an improved ViBe fast target detection algorithm is proposed to reselect and track the target continuously. |
PDF Full Text Request