The layout of urban functions is an important indicator of urban planning and development. With the rapid urbanization, the diverse human activities make urban functions highly socialized, multi-scaled and spatio-temporal dynamic. Traditional land use survey or remote sensing recognition methods are difficult to capture these characteristics of urban functions, and to reveal the gap between actual activity use and urban function planning. It is urgent to build a new framework of identifying and sensing urban function to meet the demand for real-time monitoring urban functions. The big spatiotemporal data brings a great opportunity for studying urban functions. Focusing on the identification of urban dynamic functions from the perspective of human activities, this proposal follows the logic of “spatio-temporal data-human activities-urban functions”, studies 1) the location activity semantics recognition model based on Hidden Markov Process to achieve accurate labeling the activities of massive individual trajectories; 2) the mechanism of the influence of activity spatio-temporal patterns on urban functions;3) the multi-scaled identification model of urban functions based on graph neural network and the spatio-temporal dynamic evolution mechanism to realize multi-scaled identification and highly-dynamic sensing of urban functions. The research results provide theoretical and methodological support for the systems of interacted sensing between human activity patterns and urban functions and technical support for real-time automatic monitoring of urban planning and improvement of urban governance, which have dual innovations in theory and application.