Abstract: Wi-Fi channel state information (CSI) can judge the personnel entry and exit in complex environment, which has the characteristics of low cost and high efficiency. A CSI amplitude variance threshold based-motion interval segmentation algorithm was proposed to detect the part of the CSI signal containing human motion. According to the sensitivity of different antenna and channel subcarriers to human action, an antenna-subcarrier selection algorithm was designed. According to the time difference of each antenna when the human body moved entry and exited, a human body moving direction recognition algorithm was designed. With the help of convolutional neural network (CNN), the classification and discrimination about unmanned access, one person in, one person out, two people in, two people out and two interference conditions were realized. The experimental simulation results show that the average accuracy of personnel counting can reach more than 95%.