Three-dimensional reconstruction containing texture information is a classical issue in computer vision. Considering the complexity of an indoor scene and the length of sampling image sequence captured from a random moving RGB-D sensor

conventional three-dimensional reconstruction methods suffer from limited scale and perform poor local detail reconstruction effect. This paper proposes two improvements of the RGBD-SLAM-based three-dimensional reconstruction algorithm to obtain higher quality reconstruction effect. On the one hand

the plane-primitives are incorporated as constraints to enhance robustness and accuracy of the pair-wise registration algorithm. On the other hand

to reduce the influence of RGB-D sensor large distortion

a novel exponential weight function that is motivated by a Gaussian noise model is proposed. In the experiment

the proposed method yields higher quality results compared with state-of-the-art approaches on the benchmarks dataset of the computer vision group of Stanford. Our method also achieves lower average absolute trajectory error compared with a conventional RGB-D SLAM method. Experimental results demonstrate that our method substantially increases the accuracy of camera pose estimation and quality of indoor scene three-dimensional reconstruction.