The Internet of Things realizes the connection of human and objects. Activity recognition is necessary for the interaction between information sensing devices and human. Currently, vision-based and sensor based methods are widely used, but these methods are limited in many scenes. In this paper, a new radio-frequency-based activity recognition technique was proposed, in which a few communication nodes were deployed in the monitoring area for the device-free activity recognition by analyzing the transmission packet state information. The sequential minimal optimization and K-nearest neighbor algorithms were employed for classification. The classification accuracy of walking speed of the proposed method is improved by 25.1% on average compared to the traditional method based on received signal strength indication.