Occupancy detection is essential to achieve energy efficiency in IoT-based smart buildings, so that demand-driven control of electric appliances, in particular, HVAC systems can be actuated. The IoT-based smart buildings facilitate real-time monitoring of the built environment using sensors such as temperature, light, humidity, CO2, etc. The analysis of such sensor data can detect the occupancy of the building, which in turn, is used to regulate the real-time operation of HVAC systems w.r.t human occupancy. In this regard, the existing works have analyzed smart building’s sensor data using statistical learning methods such as random forests (RF), linear discriminant analysis (LDA), and classification and regression trees (CART). However, these models suffer from bad generalization and also perform poorly on large datasets. On the other hand, the ability of deep learning (DL) to successfully learn abstract meaningful patterns from the large datasets, with better generalization position it as a candidate approach for the intelligent analysis of such IoT data. Thus, we develop a deep learning-based back propagation neural network (BPNN) model to detect the occupancy of a smart office room from ambient sensor data. The experimental results demonstrate the efficiency of our proposed BPNN model whose detection accuracy outperforms the benchmark project.