To address the high cognitive demands of augmented assembly and repair for operators, this paper presents a lightweight augmented reality framework combining object tracking and registration. Specifically, we propose a new object tracking architecture named YOLOv7-TinyMRN which is improved from YOLOv7-Tiny, reducing GFLOPS by 11.45% and model size by 6%—enabling real-time performance on resource-constrained AR devices. We further propose a 3D object registration method that integrates ROI extraction with an enhanced feature-matching algorithm based on a virtual object pose dataset to achieve a more efficient solution for pose estimation. As a result, the operated target can be recognized and tracked with the corresponding virtual model precisely overlaid, which is used to provide guidance information for the on-site operators. The experiments verify our method’s effectiveness and efficiency.