AR Long-Term Tracking Combining Multi-Attention and Template Updating

Aiming at the problem that the augmented reality system is susceptible to complex scenes and easily leads to the failure of tracking registration, a long-term augmented reality tracking algorithm combining multi-attention and template updating is proposed. Firstly, we improved the ResNet-50 network to extract richer semantic features instead of AlexNet. Secondly, the attention-based feature fusion network effectively fuses the template and search area features through a combination of dual self-attention and cross attention. Dual self-attention effectively enhances the information in the context, whereas cross attention adaptively enhanced the features of both self-attention branches. Thirdly, the ORB feature-matching algorithm is utilized to match the template and search image features, with the template updated if more than 150 matching feature points are found. Lastly, the anchor frameless mechanism is adopted in the classification and regression network, resulting in a significant reduction in the number of parameters. The results of experiments conducted on various public datasets demonstrate the algorithm’s high success rate and accuracy, as well as its robustness in complex environments.