Infant abnormal behavior detection based on improved YOLOv5 with attention mechanism

Detecting and identifying potential dangerous be-haviors in advance is a key way to predict and reduce the accidental injury of infants. Firstly, according to the behavioral characteristics of infants, a dataset of abnormal behaviors including falling, climbing and finger sucking was established. Secondly, in order to improve the effect of behavior detection and recognition, three improved YOLOv5 target detection algorithms by modifying backbone networks with attention mechanisms (YSE, YCBAM, YCoordAtt) were proposed. The experimental results showed that the proposed algorithms can improve the performance of infant abnormal behavior detection in terms of Precision, Recall, F1-score and mAP, where the mAP was increased from 76% to 82.7%.