Near misses, as a common result of unsafe behavior and/or unsafe conditions, are precursors of accidents in the construction industry. Given the complexity of near misses, seeking regularities in time series of near misses is almost impossible. We present a human dynamics model based on near-miss data to reveal common dynamic features of near-miss accidents that result from unsafe behavior of construction workers. To validate the model, we conducted an empirical case study based on unsafe behavior related to near misses from a database in the Early Warning System for Safety Risk Management in Wuhan Metro Construction, created by the Huazhong University of Science and Technology. Heavy-tailed distribution and approximate power-law distribution are observed in the inter-event time series of near misses, rather than being randomly distributed in time in current models. The strong burst and weak memory phenomena were observed in the whole, annual, and categorized inter-event time distributions of near misses. Moreover, the results showed a non-trivial, monotonous increase of the power-law exponent with the activity of different construction site in near misses, thus indicating that heterogeneity and complexity exist in general. We defined a new safety performance metric that is more reliable than the number of near misses and validated this performance metric by using expert assessment results of metro construction sites. Furthermore, we proposed further research prospects on human dynamics in construction safety behavior.