Multi-view Human Motion Analysis based on Sequential Multi-objective Particle Swarm Optimization
Yi Li · 2021
Human motion analysis from multi-view camera is a challenging problem in computer vision. In this paper, multi-view human motion analysis is formulated as a high dimensional constrained multi-objective optimization problem. A novel generative method called Sequential Multi-objective Particle Swarm Optimization algorithm (SMOPSO) is proposed for motion analysis. Firstly, nonlinear dimensionality reduction method is used to learn the low-dimensional latent space of pose state. Pose optimizing is performed in this latent space. Secondly, multi-objective optimization framework of human motion analysis is built, bi-directional silhouette based likelihood function for each camera view is designed, and MOPSO is used for pose optimization. Thirdly, in order to make MOPSO suitable for motion tracking, Sequential MOPSO is proposed by incorporating the temporal continuity information into the traditional MOPSO. Experimental results show that this method achieves accurate and stable tracking of 3D human motion.