This study proposes a hybrid tracking registration method based on saliency detection for solving the insufficient robustness of tracking registration in a complex environment and the huge feature-searching space of the augmented reality system. Based on this method, the mean-shift iteration is initially employed to predict the candidate target position. Subsequently, a two-dimensional Gaussian function is constructed with a peak at the target center, and a visual saliency map of the fusion center prior is generated. Next, the target salient feature is extracted, and the mean-shift algorithm is applied to tracking. Furthermore, the similarity measurement coefficient is used to determine whether to utilize the deformation diversity similarity-matching algorithm for relocating the target. Finally, we construct a multiscale-space fast directional binary description algorithm that performs the feature detection and matching calculation with respect to the local target area to obtain the registration parameters, and the virtual-real fusion is completed. The experimental results demonstrate that the proposed method effectively solves the problems of tracking instability and low accuracy of target detection by using the target-tracking algorithm in the cases of background clutter, target occlusion, and target loss, improving the stability and robustness of the augmented reality system.