A Multifunctional and Robust Learning Approach for Human Motion Modelling

Abstract

Human motion modelling plays an important role in the automation of the robot systems, including human motion recognition and generation. The problem is often intractable since it requires high-dimensional sequential data modelling as well as capturing both spatial and temporal correlations from motion videos. In the paper, a multifunctional and robust learning approach called broad conditional restricted Boltzmann machine (BCRBM) for human motion modelling is introduced, which employs CRBM for motion generation and broad learning system (BLS) for motion recognition. The proposed hybrid broad CRBM model has many advantages from four aspects. (1) This one-off trained model can be simultaneously employed to recognize and generate human motions with outstanding performance; (2) Without deep architecture and with the help of broad learning system, it shows out better capability in different motion generation and recognition over deep models; (3) It is more efficient as its learning process no longer need the time-consuming two-phases learning, namely pre-training and fine-tuning for deep models; (4) It is still a probabilistic graph model that has better robustness for noisy samples compared with deterministic models. These advantages are both verified and evaluated with a number of experiments, where 2D MOCAP datasets are employed.