Diverse human motion prediction based on orthogonal basis vectors

The diversity human motion prediction task predicts multiple future motion sequences from historical data. Existing research uses likelihood-based sampling, but human motion's inhomogeneity often causes modal collapse and complex training. Our paper proposes a method using a linear dynamical system to model spatiotemporal dependence, obtaining orthogonal basis vectors via Tucker decomposition. By connecting these with encoded motion residuals and sampling the Grassmann manifold with a relaxed Bernoulli distribution, we predict future motions. Compared to existing methods on Human3.6M and HumanEva-I datasets, our approach mitigates pattern collapse, improves diversity by 3%, and reduces average error by 0.07.