Visual Categorization of Children and Adult Walking Styles

Abstract

We present an approach for visual discrimination of children from adults in video using characteristic regularities present in their locomotion patterns. The framework employs computer vision to analyze correlated, scale invariant locomotion properties for classifying different styles of walking. Male and female subjects for the experiments include six children (3–5 yrs) and nine adults (30–52 yrs). For the analysis, we coordinate a minimalist point-representation of the human body with a space-time analysis of head and ankle trajectories to characterize the modality. Together the properties of relative stride length and stride frequency are shown to clearly differentiate children from adult walkers. The highly correlated log-linear relationships for the stride properties are exploited to reduce the categorization problem to a linear discrimination task. Using a trained two-class linear perceptron, we were able to achieve a correct classification rate of 93-95% on our dataset. Our approach emphasizing the natural modal behavior in human motion offers a useful and general methodology as the basis for designing efficient motion recognition systems using limited visual features.