A fall detection model based on convolutional neural network (CNN) was constructed, in order to use portable device to accurately detect the fall of the elderly and to avoid the incompleteness caused by the artificial designed features in traditional algorithms. The three-axis sensor built in the smart phone was used as the data acquisition source; the collected human body posture information was filtered, standardized, and sampled, etc. and then inputted into the designed model. Multi-layer CNN was used to train and optimize the key parameters of the model, combined with gradient descent and adaptive moment estimation optimization methods; the learned deep features were used for samples, classification. The experimental results show that the accuracy of the designed model for fall detection is significantly higher than that of the general machine learning algorithm model. In addition, the evaluation indicators have maintained a high level of stability in the detection of falls and non-falls.