In view of the lack of reliability evaluation of the existing single human pose estimation networks and the poor robustness of the deployment-oriented lightweight models, this paper proposes a testing-time-augmentation algorithm based on aleatoric uncertainty. Firstly, the diverse outputs are obtained through stochastic parallel data augmentation and model inference. Then, the reliabilities are acquired by calculating the aleatoric uncertainty of those outputs. Finally, the outputs are fused according to reliabilities and a more accurate and robust result with evaluation is finally got. Experiments show that the algorithm can be a plug-and-play applicable to the existing single-human pose estimation models. By using this algorithm, a more accurate and robust result along with uncertainty can be got than before.