Collective intelligence of temporal statistics for segmenting sustained infant feeding behaviors in Videos
Xinpeng L. Liao, Chengcui Zhang, Wei-Bang Chen, Paula C. Chandler-Laney · 2016
Converging evidence suggests that the risk of obesity may be at least partially determined during infancy. Instead of measuring total caloric intake, obesity research shifts its paradigm to examining infant feeding behaviors so that the risk of overeating can be detected in early life. To the best of our knowledge, this paper for the first time proposes a novel video analysis protocol with which to discover meaningful, objective, and valid estimates of infant feeding patterns, e.g., the sucking counts over a sustained feeding timeframe as an initial attempt. The underlying premise is that, after tracking a collection of visual features, a back-and-forth periodic movement exhibiting bimodality properties can be segmented from other cluttered ones to indicate sucking behavior, irrespective of camera angles and subjects. We introduce and formulate this motion-of-interest segmentation as a Linear Programming (LP) problem where the optimal solution intelligently infers the sucking count from a collection of temporal statistics after feature tracking. Preliminary experiments on video sequences of different camera angles and subjects successfully validate our hypothesis and show promising results of the solution.