An efficient heuristic method for infant in/out of bed detection using video-derived motion estimates

Camera-based infant monitoring has received substantial attention because of its unobtrusiveness, allowing long-term and continuous monitoring. To assess infant sleep, it is required to know whether the infant is in bed or out of bed. This can automate monitoring for analyzing data solely when the infant is in bed, providing accurate assessment of sleep quality. In this work, we propose a method aimed at detecting infant in/out of bed status for a 24 h period using motion estimates derived from an infrared video camera. The method is based on several heuristic decision-making rules to examine the motions between peaky motions that are expected to express activities or events of putting the infant in bed or taking the infant out of bed. For example, when these events are identified, intuitively, if a period between two events has consistently motions (caused by infant's body movements), this period should be detected as infant 'in bed', otherwise as 'out of bed'. Our proposed method would largely increase the efficiency (i.e. reduce the computational load) compared to some advanced computer vision algorithms that try to recognize infant's face or torso. To separate the motions caused by infant body movements and other activities from, for example, parents, a region of interest (ROI) is configured. This ROI can be fixed and customized. 77 datasets of 24 h recordings from five infant participants were analyzed in this study. Experiments were conducted in a free-living environment in the participants' own home. Results show that an average accuracy of 95.8% using a fixed ROI configuration, and that of 96.9% using a customized configuration were achieved. This indicates that our proposed method is reliable for infant in/out of bed detection.