This paper investigates and compares machine learning models for classifying gender and ethnicity with human anthropometric measurements as input attributes. Optimal attribute sets are identified through individual measurement ranking and subset selection. These sets are further down-selected by taking into consideration the acquirability of measurements. Using the Civilian American and European Surface Anthropometry Resource (CAESAR) database as training and test datasets, the investigation has achieved a classification rate over 96% for gender (male/female) and 80% for ethnicity (White American/African American), respectively. Furthermore, the effect of random measurement noise on the classification performance is investigated to find a preliminary performance boundary for the classifiers. This study shows that gender can be predicted with high confidence and robustness from a few torso dimensions, while ethnicity can only be estimated roughly from limb dimensions under real-world conditions. The approach developed in this paper can be used with other image analysis software to achieve better understanding of human attributes contained in video imagery and to facilitate automated content analysis and decision making.