Effective recognition of human actions is a must for the further development of artificial intelligence. In this paper, wearable sensors are utilized to collect human action data based on time series. Combined with median filtering technique to process the action data and keep the edge information, instantaneous and local features are highlighted by wavelet transform. Long and short-term memory network (LSTM) is added into the convolutional neural network to solve the limitation of insufficient dependency of the convolutional neural network in learning human actions and improve the model classification accuracy. A large human action database is selected for model training and testing to compare the performance advantages of this paper’s model. The results show that the actual number of misclassifications of this paper’s model in the 2 datasets is only 16 and 5. During the training process, the loss value is always no more than 0.1, and it shows a stable decreasing trend. The average recognition accuracy is steadily improved from 0.03891 to 0.76787. In the multi-method comparison, the accuracy of CS and CV metrics of this paper’s model is 87.39% and 90.87%, and that of Top-1 and Top-5 metrics is 41.76% and 60.63%, which is higher than that of the comparison methods. The model in this paper can effectively realize the smooth recognition of human body movements with high accuracy.
