In order to achieve real-time fall detection of special people in the video and reduce the false detection rate. By using the improved pose estimation network to extract the human body key points, the changes of the frame key points before and after were studied to track the human body and detect the fall behavior. In order to achieve the real-time effect of the pose estimation network on the embedded platform, a lightweight structure with attention mechanism was used to build a deep convolution network to extract the joint coordinates of the human body and synthesize the complete skeleton information. The results show that the accuracy of the pose estimation algorithm with attention mechanism is improved on different datasets. At the same time, the real-time fall detection is achieved with low error detection rate on the embedded platform. It can be seen that the improved pose estimation algorithm and the nodal judgment method can better realize the fall detection of human body.