Deep Learning-Based Infant Behavior Feature Classification Method and System

Abstract

The present invention provides a method and system for classifying infant behavior characteristics based on deep learning, including: Step 1: Produce an infant video dataset, and classify and annotate time segments for the videos in the dataset; Step 2: Produce an infant key point dataset; Step 3: Preprocess the human key points of the infant video dataset; Step 4: Use a deep learning model to build an infant behavior classification model, send the human key point features of the infant video dataset into the infant behavior classification model, calculate the classification probability output by the model, the total loss between the output time segment and the label, and use the total loss for backpropagation to update the model parameters. After the model parameters converge, the training is completed; Step 5: Judge whether there is a dyskinesia in the video according to the classification probability of the classification branch, and obtain the time segment when there is a dyskinesia in the video. The present invention can detect whether an infant lacks dyskinesia and relieve the burden on doctors of having to watch infant movement videos for a long time.

Images