Falls are a major cause of disability and even death in the elderly, and fall detection can effectively reduce the damage. Compared with cameras and wearable sensors, Wi-Fi devices can protect user privacy and are inexpensive and easy to deploy. Wi-Fi devices sense user activity by analyzing the channel state information (CSI) of the received signal, which makes fall detection possible. We propose a fall detection system based on commercial Wi-Fi devices which achieves good performance. In the feature extraction stage, we select the discrete wavelet transform (DWT) spectrum as the feature for activity classification, which can balance the temporal and spatial resolution. In the feature classification stage, we design a deep learning model based on convolutional neural networks, which has better performance compared with other traditional machine learning models. Experimental results show our work achieves a false alarm rate of 4.8% and a missed alarm rate of 1.9%.