Augmented Reality Long-term Tracking Based on Multi-Attention
Guo Mengru, Qiang Chen · 2022
Aiming at the problem that the augmented reality system is susceptible to complex scenes leading to the failure of tracking registration, a long-term augmented reality tracking algorithm combining attention mechanism is proposed. Firstly, the improved ResNet-50 network is used to replace AlexNet network to extract richer semantic features. Secondly, the attention-based feature fusion network effectively fuses template and search area features. The network is consisted of the dual self-attention module and the cross-attention module. The dual self-attention module effectively enhances the context information, and the cross-attention module adaptively enhances the features of the two self-attention branches. The anchorfree mechanism is introduced in classification and regression network, which greatly reduces the number of parameters. Experimental results show that the algorithm achieves high success rate and accuracy on various public datasets, and also shows good robustness in complex environments.