Present detection can remotely detect whether a person appears in specific scene. Channel state information (CSI) can provide high precision environment detection by its characteristic, and this technology can often be adopted for indoor localization or presence detection nowadays. We propose a device-free multiple presence detection system by using particular preprocessing method for our system, and a two-stage learning system combining convolutional denoising autoencoder (CDAE) and neural network (NN) to classify different cases of presence detection. The first stage contains several one-dimensional convolutional hidden layers which can denoise and reduce the dimension of data. The second stage classifies the cases for the purpose of presence detection. The proposed system can detect the presence of multiple persons at hotspots with high estimation accuracy.