Vision based Abnormal Action Detection of Children in Wards

In clinical care, hospitalized children are prone to get secondary injury such as falling from the bed. Avoiding secondary injury to patients has become the core indicator of medical care quality evaluation. Falls are the most common and most dangerous behavior of children in hospital. And vision based abnormal action detection of children in wards is very challenging because of the unnoticeable appearance and the complex ward environment. Previous approaches achieve anomaly action detection in urban and industry by the skeleton of human. However, they cannot effectively detect and alarm abnormal actions of hospitalized children in the ward, because their model or Kinect cameras cannot obtain the accurate skeleton of the child in real time. In this paper, we propose an intelligent video surveillance system aimed at detecting and alarming abnormal actions of children in wards. Experimental results demonstrate that our system achieve 85% accuracy in detecting and alarming abnormal actions of children in wards, and our system can process information from two beds simultaneously with a speed of 12ms per frame. The daily video data of children in the ward is also saved locally, which is exceedingly useful for augmenting risk action datasets of hospitalized children. In the future, we will improve our proposed system by obtaining state information of children, thereby reducing the false positive rate of the system.