Two three-dimensional tracking registration methods combined with Riemannian manifold object constraints are proposed to solve problems of low accuracy and instability of three-dimensional tracking registration in sparse and complex scenes. A deep convolution neural network is used to extract three-dimensional instance objects from the location by analyzing reasons that affect registration accuracy in sparse and complex scenes. The three-dimensional tracking registration model is established according to the Riemannian manifold constraint relationship of instance objects in different states. The stability of the three-dimensional tracking registration algorithm is improved by combining inertial sensors, and cumulative error is optimized using instance object labels to improve algorithm robustness. The proposed algorithm can effectively improve the accuracy of three-dimensional tracking registration. It can improve the performance of augmented reality systems and be applied to power system navigation, medical, and other fields.