Studying human behavior from infancy: On the acquisition of infant postural data

The study of human behavior using infants as target subjects is very attractive since these individuals are minimally affected by cultural background and display the fastest rates of evolving cognition and physique, opening possibilities to longitudinal but relatively short-term research. Naturally, important customers of infant movement data are healthcare practitioners and scientists at the cutting edge of the understanding of human development and related disorders, in particular Autism Spectrum Disorder (ASD). Here we provide evidence that, as opposed to the current practice, these studies demand non-invasive instrumentation to measure movement, so the right paradigm to obtain the data will most likely depend on computer vision based pose estimation. By surveying the interdisciplinary literature on infant motion capture, we show that, up to now, very little has been done to consider infant data in vision, and no method has treated the problem of measuring infant movement as special problem of its own, but rather a special case of the general human movement capture problem. We oppose this position and propose the use of canonical postures as an implementation of the principle of stability noted by developmental psychologists, and exemplify how these postures and age-related data could be used to potentially improve existing pose estimation systems. We also show preliminary results suggesting that canonical postures may be recognized using global, low-level contour features augmented by mid-level features like elongatedeness; these results are consistent with previous work in infant pose estimation using pressure-based sensors.