In this paper, we investigate methods based on Convolutional Neural Networks (CNN) and online learning for infant body part segmentation in videos. Long videos of infant movement can be used to detect abnormal neurological development caused by various reasons. In order to perform movement analysis and assist with diagnosis, it is crucial to accurately segment infant body parts in videos. Existing CNN based segmentation methods can perform reasonably well given sufficient training data, but cannot reach an accuracy level that is needed for clinical use. We propose and experimentally validate three methods that significantly improve segmentation accuracy and efficiency: (1) content-dependent training sample selection reduces the amount of training data needed while boosts segmentation accuracy, (2) a weighted loss function with different weights for different body parts is shown to both improve accuracy and reduce computation time, (3) online learning by dynamically augmenting the training set with a small number of sample frames from new videos significantly improves testing accuracy on the new videos. Our experiments are conducted on infant movement videos that exhibit large variations in object pose, shape, appearance, background and camera viewpoints. Both qualitative segmentation results and quantitative comparisons demonstrate the superiority and efficiency of our proposed approach.