With the growth of informatization and the rapid development of technologies such as the Internet, big data, and the 5G communication, the connection between the information world and the real world is getting closer and closer. As a technology that integrates the virtual world with the real world, augmented reality is being widely used in many fields such as industrial manufacturing, medical care, cultural tourism, and interactive entertainment. The 6-degree-of-freedom pose tracking of known objects is the core part of augmented reality, and only when the pose of an object is correctly acquired can the virtual information be further fused on the object. As augmented reality-related technologies continue to mature and users wish to further expand the application scenarios, higher requirements are placed on the accuracy and robustness of augmented reality systems, and object tracking technologies are facing new challenges. Traditional object tracking methods use hand-designed features for matching, which are difficult to cope with complex application environments such as weak textures and occlusions, and the accuracy of tracking is also limited. In order to obtain higher accuracy and robustness, deep learning methods have become popular in recent years. However, it is computationally intensive and usually difficult to be applied to mobile augmented reality devices. In addition, the large amount of time and data required for training is also a major obstacle to its application.
To address these problems, this paper will study the object tracking algorithm based on deep learning algorithm, which can meet the computational volume of augmented reality applications, and the computational process does not rely on the prior knowledge of the object model, so that it can have a better tracking effect on the objects outside the training. The main research of this paper is to investigate a deep learning based object tracking algorithm, which on the one hand, enhances the ability of the algorithm to resist occlusion by analyzing the depth projection map of the object model; on the other hand, the structure of the algorithm is designed and trained in such a way that the prediction results do not depend on the prior knowledge of the object model to be tracked, and has a more general feature description capability, so that it can also perform better on objects outside of training. This allows the algorithm to perform better on objects outside of training. Finally, in order to enable the algorithm to have a smaller computational effort without affecting its accuracy, this paper also proposes an image object interception method that recovers the field of view by camera parameters, making the input of the algorithm more accurate and concise.
To verify the effectiveness of the proposed method, experiments are conducted on three classical datasets in object tracking and compared with the current mainstream methods. The results show that among the selected control methods, the proposed method achieves the highest average accuracy on two datasets and can effectively cope with complex application scenarios such as occlusion. To verify the feasibility of applying this paper's algorithm to augmented reality application scenarios, this paper is also tested on mobile devices and successfully tracked objects in real environments, demonstrating the great potential of applying this paper's algorithm to mobile.
|Keyword||物体追踪 增强现实 深度学习 Transformer 迁移学习|
|Sub direction classification||计算机图形学与虚拟现实|
|planning direction of the national heavy laboratory||其他|
|曹恩源. 增强现实中物体追踪关键技术研究[D]. 模识楼201. 中国科学院自动化研究所,2022.|
|Files in This Item:|
|Recommend this item|
|Export to Endnote|
|Similar articles in Google Scholar|
|Similar articles in Baidu academic|
|Similar articles in Bing Scholar|
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.