With the continuous development of science and technology, human activity recognition has gradually become an indispensable part of our lives. In indoor positioning, emergency rescue, medical security, automatic driving, mine exploration and even military fields, people put forward higher and higher requirements for human activity recognition technology. At the present time, the activity recognition methods developed by researchers are divided into three categories: human activity recognition methods based on vision, environmental sensors and human body sensors. However, human activity recognition methods based on vision are easily affected by the environment and easily cause privacy problems. The method of hu-man identification based on environmental sensors relies too much on the installation of infrastructure and can only work in a limited area; The human body recognition method based on pedestrian wearing sensor needs high cost. According to the data collected by smart phones, our paper pro-poses a human activity recognition method based on Stacked Denoising Self-Encoder (SDAE) and XGBoost. The data collected by the smart phone sensor is denoised by using SDAE, and features are extracted by unsupervised learning. As a gradient lifting framework, XGBoost can effectively extract and analyze the key features in activity data, build an efficient gradient lifting decision tree model, and realize the accurate identification and classification of human activity. We apply our method to open data set UCI HAR and compare it with the traditional algorithm. The experimental results show that the performance index of our proposed method is better than the traditional algorithm.