Data availability often becomes a major hindrance to the development of human centric, computer-vision based technologies, as a large amount of data is usually required for algorithm training and validation, especially when deep learning is used to develop algorithms. Synthetic data which are produced by modeling and simulation could be used to expand and/or supplement real world data which are otherwise not available. While human activity modeling and simulation has achieved success in creating synthetic environments for simulation based training and virtual reality, whether it can be used to generate synthetic data which satisfy requirements for machine learning is yet to be proved. In this paper, using human activity modeling and simulation to generate synthetic human activity data for machine learning is investigated. The needs for synthetic data are identified from the perspective of human centric, computer-vision based technology development. The basic requirements of synthetic data are defined in light of machine learning. Factors that contribute to the fidelity and applicability of synthetic data are analyzed. In particular, two factors related to human activity modeling and simulation, bio-fidelity and variability are investigated. Several modeling and simulation tools and game engines (e.g., 3dsMAX, Unity, and NVIG) are used for data generation, and their performances are compared and evaluated. Synthetic full motion videos are generated in electric-optical and infrared modes and tested by machine learning algorithms. The testing results along with examples of synthetic imagery are illustrated in the paper.