Human motion prediction based on spatial information and channel attention

With the development of deep learning and computer vision technologies, human behavior prediction has gradually become a research area of great interest. Existing work has introduced the Transformer architecture into human behavior prediction, but due to the differences in data types and dimensions, one-dimensional position coding is difficult to effectively capture the spatial correlation between human joints. At the same time, there are problems such as poor differentiation of different channels and cumbersome training process. In order to solve the above problems, this paper is based on the denoising diffusion model and combines with the Transformer architecture to predict human behavior. The method gradually adds noise to the skeleton sequence in the forward process, and encodes the position of the joints by projecting the 3D human skeleton coordinates to the 2D position plane in the reverse denoising process, so as to effectively capture the relative positions and distances of the joints in the space. Then the compression and excitation module learns the interdependence between channels to dynamically assign channel weights to enhance the model's ability to distinguish between different channels. The experimental results show that the proposed method has advantages in prediction accuracy and stability compared with the existing methods, and the experiments on the Human3.6M dataset and HumanEva-I dataset verify the effectiveness of the method.