Infant Gait Modifications: Integration of Computer Vision with Human Annotation Provides Accurate Step Classification and Location as Infants Navigate Varied Terrain

Functional locomotion requires perception of affordances—the fit between body and environment that makes particular actions possible. To cope with changing affordances while walking over varied terrain, walkers must modify their steps as they approach and navigate each ground surface. Newly walking infants have the physical wherewithal to modify their gait by slowing down and taking shorter steps, but they do not do so systematically or prospectively. To test how gait modifications develop, we video recorded new and experienced infant walkers as they approached and crossed slopes and bridges. The long-term aim is to measure whether, when, and how infants modify their gait (relative to slope degree and bridge width, the location of the obstacle in space, and typical gait on flat, wide surfaces). Useful data require high precision in classifying steps and identifying the 3D location of infants' feet at each moment (despite the idiosyncrasies of infant movements and frequent occlusion and motion blur)—requirements beyond the capabilities of human annotation or computer vision alone. Thus, we built an integrated human-machine system to identify each step and its XYZ coordinates as infants approached and crossed the obstacles. We capitalized on the ability of human annotators to classify infants' movements into steps and of computer vision to identify the 3D coordinates of the feet in each video frame. We demonstrate the feasibility of this integrated, human-machine system to investigate the development of prospective infant gait modifications.