A Bayesian Dynamical Approach for Human Action Recognition

We introduce a generative Bayesian switching dynamical model for action recognition in 3D skeletal data. Our encodes highly correlated data into few sets of low-dimensional temporal processes and from there decodes to the motion their associated labels. parameterize these with regard deep autoregressive prior accommodate both multimodal higher-order nonlinear inter-dependencies. This results latent that parses meaningful intrinsic states dynamics enables recognition. These sequences provide visual quantitative interpretations about primitives gave rise each class, which have not been explored previously. In contrast previous works, often overlook dynamics, our method explicitly transitions is generative. experiments on two large-scale datasets substantiate superior performance comparison state-of-the-art methods. Specifically, achieved 6.3% higher classification accuracy (by incorporating framework), 3.5% better predictive error employing second-order transition model) when compared best-performing competitors.