The occupancy information is essential for ensuring indoor comfort and enhancing building energy efficiency. Through image-based object detection, the indoor occupancy information can be obtained. However, small targets of the human always reveal the weak feature expression ability in indoor scenes, which increases the difficulty of detection. In addition, due to the complexity of the indoor environment, the possibility of false detection of objects similar to human characteristics is increased. Aim at the above problems, a new indoor occupancy detection model by integrating the efficient multi-scale attention mechanism into the EfficientDet model is proposed herein. Firstly, to solve the problem of indoor human small target detection, the EfficientDet model is innovatively applied to indoor human detection. Secondly, for the purpose of reducing the interference from background objects and focus on human key information, this paper infuses the EMA mechanism with the EfficientDet model. Finally, a self-built classroom dataset has been constructed for validating the occupancy detection performance of the model. The experimental results show that the average recognition accuracy of this method on public dataset SCUT-HEAD-Dataset Part B is 92.85%, and that on self-built classroom dataset is 96.15%. Compared with other mainstream algorithms, EMA-EfficientDet also shows good performance.