To improve the accuracy of pedestrian gait recognition, a real-time recognition method for seven types of daily gait based on time domain features and a convolutional neural network is proposed. First, the acceleration and angular velocity sensors of the mobile phone are used to collect the time series data of the sensor on the x-, y-, and z-axes in the states of sitting, standing, walking, jogging, and squatting. The convolutional neural network is used for offline recognition, and the sampling times (t = 0.5, 1, and 1.5 s) are used for online real-time recognition. Then, two types of gait, ascending and descending stairs, are added for offline and online recognition and compared with the previous five types of classification recognition. We concluded that the gaits of walking and ascending stairs, as well as those of standing and descending stairs, are similar, which lead to the decrease in overall classification accuracy. Therefore, the time domain features of the total value of the acceleration sensor on the x-axis, the maximum change on the z-axis, and the change of steps during the sampling time are extracted, and the convolutional neural network model and time domain features are combined for online recognition. The experimental results showed that this method can significantly improve the transfer rate of gait recognition information and provide a new idea for gait recognition in the fields of motion detection and elderly monitoring.