3D human pose estimation is eagerly in demand as pose estimation is widely used in real life, such as human recognition, action detection, and robot imitation learning. However, the previous work on 3D human pose estimation is often paired with average errors on the human joint position which gives each joint the same weight. This loss function ignores that the estimation errors of different joints have different impacts on the pose estimation. To tackle this problem, we propose a linked temporal convolution model with a weighted joint loss function to estimate 3D human pose in videos. In this model, we start with predicted 2D key-point then estimate 3D poses by linked temporal convolutions network which link hierarchical features from different layers and group features together to increase the performance. Through experiments, we find our network achieves more precise performance on short-term prediction.