Motion Diffusion Model for Long Motion Generation

Long-term human movement prediction has become a key research focus, driven by its numerous applications in domain such as robotics, virtual reality, video analysis. Traditional methodologies, primarily based on artificial modeling, often unable capture the complicated dynamics of human movement. Recent developments have explored recurrent neural networks and generative models for sequence prediction, but challenges persist in achieving accurate and realistic long-term predictions. This study introduces a deep generative motion infiller leveraging a diffusion model, capable of seamlessly generating motion sequences of arbitrary length from historical or sparsely available key-frames. This model adeptly addresses the complexities of filling in missing or occluded body movements by utilizing observed motion data and predicting new sequences. The proposed method incorporates a Transformer-based structure optimized for lightweight operation, which is employed within a sliding window framework to establish a recurrent generation system. Additionally, a novel method is proposed for style transfer, applicable during inference, enhancing the diversity and the resulting sense of motion. The proposed model also demonstrates effectiveness in generating short-term and long predictions, comprehensive evaluations indicate that the proposed approach yields more diverse and realistic motion predictions compared to existing methods, we evaluated the model in indicate of global and local diversity, which have the better result.