In order to improve the recognition accuracy of traditional fall detection system and reduce the recognition time, a new fall detection model is proposed. The skeleton node obtained by depth vision sensor of Kinect V2 is used as the sample data source, and the improved k-means algorithm is used to calculate the clustering center point, and on this basis, the fall detection feature data is extracted. After reconstituting the feature data into 5×5 training sample data, the convolution neural network model is input to train and learn, and the optimal fall detection model is obtained. Experimental results show that the new detection model has higher recognition accuracy and faster operation speed than the traditional fall algorithm, which guarantees the real-time and robust requirements of the system.