Child abuse is a grave and pervasive social problem with profound consequences for both individual victims and society as a whole. Proper detection is important not only in recognizing child abuse but also in applying appropriate penalties for those who perform violence on children. In this paper, YOLOv5 and YOLOv8 are used to build models detecting whether there is any child violence performed in surveillance videos. The child abuse behaviors examined in this study were pinching, kicking, slapping, and choking. Experimental results on the dataset of 12666 images related to these four behaviors and extracted from videos showed that the model built by YOLOv8 is better than the other. It obtained the IoU measure of 80.9% and the F1-score measure of 97%.