Robust 3D Human Pose Estimation Model Based on Temporal Convolution

摘要

The existing temporal convolution model does not take full advantage of the characteristics of the low level network, and there is noise in the input 2D joint sequence. Therefore, a temporal convolution model based on intermediate supervision is proposed to estimate 3D human posture. Firstly, a 2D joint filter is designed to locate and correct obvious noise points, so that the input data is closer to the real data. Secondly, an intermediate supervision structure is added to the sequential convolution model to make full use of the characteristics of the lower level. Finally, the slicing function of the original model is replaced by the pooling layer to improve the abandonment of some features of the residual layer. The experimental results in the Human3.6M show that the model with 2D joint filter and network optimization is superior to the existing temporal convolution model, and the prediction error is reduced by 2.56% compared with the original model.