An integral part of visualizing an air-conditioned space is to know its occupancy in real-time, in order to make intelligent control decisions about the operation of its Air Conditioning and Mechanical Ventilation (ACMV) system. The sensing mechanisms used in occupancy estimation such as cameras and wearable sensors are generally intrusive and expensive. Alternatively, the effect that occupants have on environmental parameters such as CO2, temperature, humidity and pressure can be utilized to extract information about the occupancy levels. Environmental sensors are relatively inexpensive and are non-intrusive. From these sensor data, we need to extract and select relevant features that may yield occupancy information. The filter model feature selection approach used in previous works compromises on the classification accuracy in order to limit the computational burden. An alternative is the wrapper model of feature selection, which uses the inference algorithm itself to search for the best features. It guarantees better classification accuracy but is computationally expensive, especially with slow iterative machine learning techniques such as the Artificial Neural Network (ANN) used in previous works. To address this problem, this work capitalizes on the fast learning speed of Extreme Learning Machines (ELM) to implement a wrapper model of feature selection. To the best of our knowledge, the use of the wrapper model in an occupancy estimation problem has not been documented. A comparison between the filter and wrapper model feature selection is made. The tracking accuracy was seen to have notably improved with the wrapper model. Also, it was demonstrated that the pressure data, which has not been used for occupancy estimation in previous works, is useful.