PT - JOURNAL ARTICLE AU - Yamamoto, Hiroki AU - Kitamura, Koji AU - Yamanaka, Tatsuhiro AU - Nishida, Yoshifumi AU - Mizoguchi, Hiroshi TI - PW 2646 Dynamic prevention of falls from balconies using artificial intelligence cameras AID - 10.1136/injuryprevention-2018-safety.621 DP - 2018 Nov 01 TA - Injury Prevention PG - A225--A225 VI - 24 IP - Suppl 2 4099 - http://injuryprevention.bmj.com/content/24/Suppl_2/A225.1.short 4100 - http://injuryprevention.bmj.com/content/24/Suppl_2/A225.1.full SO - Inj Prev2018 Nov 01; 24 AB - Among childhood accidents that result in death, falls constitute the highest proportion. In Tokyo alone, in the five years from 2011 to 2015, there were 114 emergency transports of children under the age of 5 to medical institutions that involved falling from the windows or balconies of houses and apartments.In the present paper, we develop a dynamic safety system for preventing falls from balconies using artificial intelligence technology for recognizing the behavior of children and environmental risk on a balcony using RGB-D cameras.In order to help prevent balcony falls, the developed system has two functions: (1) the ability to detect a single child without an accompanying adult and (2) the ability to calculate fall risk based on shape evaluation of items present in the balcony environment. In order to verify the developed solitary child detection and fall risk calculation functions, an experiment was conducted at a mockup living laboratory designed to simulate a general household environment. We also conducted an experiment to determine whether the developed system could operate even under the influences of direct sunlight, wind, and electric lighting.For the solitary child detection function, we confirmed that the proposed system can estimate the age of a child to within an error of 5% in terms of years of age (one to six years of age). For the fall risk calculation function, we compared the estimation values of the proposed system to true values obtained each time that we changed objects and placement positions, and the error was within 5%. These two experiments confirmed the effectiveness of the developed functions.Future tasks will include expanding the detection function to include other dangerous shapes, such as obstacles that would be easy for a child to climb.