 |
Child'sPlay Smart Toys
Children spend a large amount of their time playing with toys. As such, we feel that toys may be leveraged to assist in the automatic capture and annotation of developmental data. We have developed six toys with embedded sensors for the wireless transmission of motion, sound, and touch information during play. These sensing modalities can be useful in collecting information on components of developmental milestones such as baby babbles, grip strength, and physical manipulation abilities. Our toys include a plush puppy rattle, a plush caterpillar, plastic Lego QuatroTM compatible blocks, a plastic stacking ring toy, an abstract shape resembling a cooking pot lid, and plastic stacking domes. The plush toys contain social faces to encourage engagement with the toys. The ring, lid, and dome toys are rounded objects based on a similar circumference to encourage stacking, covering, and scooping activities. Each toy continuously transmits data via the Bluetooth sensor to a mobile computing device where it records and processes the data. To detect the play activities we use the adapted approach that classifies activities based on the boosting of aggregate features computed over a short temporal windows.
We have conducted pilot plays tests with five adults and three children over a minimum of two 10 minute sessions per participant. These sessions allowed us to test the durability of the toy designs, the appeal of the toys, the data transmission capability of the toys, and the ability of our algorithm to properly detect relevant play activities. Our initial exploration of the play data has revealed 25 distinct classes that comprise the basic play actions that our system will recognize. As a result of our initial work, we learned that our ABS plastic and conductive textile toys were durable, functional as toys, used by the children, and concealed the sensor from the participants. Our toys withstood throw, drops, and kicks that occurred during play. However, our play tests did expose an important challenge with constructing the automatic recognition portion of our system. The most prominent challenge will be collecting enough training examples to build models for recognition. While we can script play scenarios to help encourage children to engage in the types of play we are trying to detect, there is no guarantee that the child will be willing to comply. This problem is further complicated by the fact that our target age group is children ages 10-24 months old. Some of these children may not have yet formed language, nor will they be receptive to instructions by adults. Thus, we cannot instruct them to play with the toys in a way that will elicit proper training data.
More details can be found in our workshop publication http://www.cc.gatech.edu/~turtle/my_papers/westeyn_IDC2008.pdf
This work is the basis of Tracy Westeyn's thesis work. More details can be found
Last modified 2 September 2008 at 11:07 pm by turtle |
