Average travel speed for road sections is important information for traffic condition evaluation, transportation management, and so on. Vehicular crowdsensing can be used to collect data including real-time locations and velocities of vehicles, which can be further utilized to estimate travel speed. However, due to the uneven spatial-temporal distribution of vehicles and the variation of data-offering intervals, vehicular crowdsensing data is usually coarse-grained. To handle the coarseness problem as well as estimate travel speed accurately, this work proposes an approach named STC (abbreviation of Spatial-Temporal Correlation) to produce a fully covered distribution of travel speeds for road sections in both time and space domains based on vehicular crowdsensing data. STC exploits the spatial-temporal correlation among travel speeds of road sections by introducing a time-lagged cross correlation function. We use a time lagging factor to quantify the time consumption of vehicles travelling along road sections, and determine the time lagging factor self-adaptively by tracking the locations of vehicles at different road sections. Then, utilizing the local stationarity of the cross correlation, we reduce the problem of finding the missing travel speed for a single road to a minimization problem. Finally, we fill all the missing travel speeds in a recursive way by using the geometric structure of a road network. Experiments based on real taxi trace data show that STC has a better estimation accuracy, in comparison with typical methods in use.