Abstract: Extraordinary load is an essential part of the live load on buildings. However, the traditional model based on load cells makes the field survey be difficult, and the lack of survey data has led to the long-term stagnation of extraordinary load studies. This study proposes a decoupled stochastic model for extraordinary load, which decouples extraordinary load into two independent random variables of load quantity and magnitude. The proposed model enables unified modeling of different extraordinary load events and simplifies the load investigation by conducting sample statistics separately. Based on the decoupled stochastic model, a machine vision survey method, combining object detection and multiple-object tracking, is further proposed to conduct extraordinary load surveys sustainably using widespread public surveillance facilities. Taking the extraordinary load of crowd gathering as an example, the implementation of the machine vision survey method is presented, and the effectiveness of the new survey method is verified by the survey experiment under actual scenes. The decoupled stochastic model combined with the machine vision survey method provides a feasible way to obtain large samples of extraordinary load and study the design value of floor live loads.