We address the problem of estimating the occupancy levelsin rooms using the information available in standardHVAC systems. Instead of employing dedicated devices, weexploit the significant statistical correlations between the occupancylevels and the CO2 concentration, room temperature,and ventilation actuation signals in order to identify adynamic model. The building occupancy estimation problemis formulated as a regularized deconvolution problem, wherethe estimated occupancy is the input that, when injected intothe identified model, best explains the currently measuredCO2 levels. Since occupancy levels are piecewise constant,the zero norm of occupancy is plugged into the cost functionto penalize non-piecewise constant inputs. The problemthen is seen as a particular case of fused-lasso estimator byrelaxing the zero norm into the `1 norm. We propose bothonline and offline estimators; the latter is shown to performfavorably compared to other data-based building occupancyestimators. Results on a real testbed show that the MSE ofthe proposed scheme, trained on a one-week-long dataset, is half the MSE of equivalent Neural Network (NN) or SupportVector Machine (SVM) estimation strategies.