Crime against children has been a big social concern. A measure to protect children from such crimes lets them hold smartphone so as to keep in touch with their guardians in any situation. An incident may happen in an unexpected manner which does not allow children to get helps with their phone. Some special measure is needed to handle potentially threatening situations.
Children usually show routine life patterns: Their activities are conducted in a small number of areas such as home, school, and extra-curricular activity academies. Their phones are intentionally turned off from time to time to concentrate on some activities. Deliberate analysis of child’s behaviors of data obtained by her/his phone can tell whether she/he is in potentially unsafe situation.
This thesis presents a child care service that uses contextual data acquired with smartphone. The service collects sensor data such as GPS data, time, and battery charge status data. To model child commuting patterns, we propose a dynamic Bayesian network model which is trained by the collected sensor data. The proposed dynamic Bayesian network model represents child commuting patterns and smartphone turn-off patterns with probabilistic distributions encoded in a probabilistic graphical model. With a smoothing filter and a Kalman Filter, the proposed method in the training phase filters the GPS data of child commuting trajectories to reduce noises. For fast inference of unknown quantities, Rao-Blackwellized particle filter is used which allows an approximate fast inference. Context-aware safety actions are defined based on the inferred values.
A real data set had been collected for a middle school girl student over 3 months. In the experiments, the data set was used to train the dynamic Bayesian network and to test the model. The proposed model successfully learned the patterns for the training set and could detect abnormal situations. We have implemented a prototype system for the child safety care service and observed that the service works as designed.