Image banner for technology event detection section

Event Detection

  In order for the room to provide interesting, reactive feedback, it must be able to detect "events."  

What is an Event?

At certain times in the story, the control program needs to know what the children are doing. In the KidsRoom, the control program uses the sensor outputs to compute yes-no answers to several questions:

  • Is everyone in a group?

  • Is everyone on the bed?

  • Is everyone on the path?

  • Is everyone moving around the path or standing still?

  • Have the kids just screamed?

  • Is someone near a particular object?

Event Sensors and Detectors

The KidsRoom uses simple rules to compute answers to the above questions. For example, the "in-a-group" detector gets the position of each person from the vision tracker and checks to make sure that every person is within some pre-determined distance of another person. The "on-the-bed" detector fires true when the tracker blobs for each person have merged into a single large blob which includes the blob for the bed (the bed can be moved and therefore must be tracked). The "just-screamed" detector checks if an amplitude threshold has been achieved over a short temporal window.

Difficulties

One problem we encountered when designing the KidsRoom was that "simple" events are strongly context-sensitive. For example, our "in-a-group" detector will signal false continuously if one mischievous kid refuses to cooperate with the remaining children. In this case, a more robust detector might ignore the "outlier" given that he hasn't been following the rules for a while (e.g.using a "bad-kid" detector). Similarly, if one child is scared and remains on the bed while other children explore the forest on the path, the "in-a-group" detector should ignore this child as well. These cases are not handled in the current system, but offer an interesting, challenging area of future research in action and event understanding.

  Story - Playspace - Technology - People - Info  

The KidsRoom - Perceptual Computing Group - MIT Media Laboratory