چكيده به لاتين
The Abnormal Detection or Anomaly Detection in general, is the problem of searching the data space to find different and unusual instances. These kind of instances happen rarely but may include valuable information. The problem of abnormal detection is used in various fields, such as Human Fall Detection or Fall Detection in short. An accidental fall may lead to serious damages, especially in the case of elderlies. Sometimes, fell person can’t standup again by himself and will be at the risk of long lying consequent injuries. The goal of a fall detector system is to distinguish well between a usual daily activity and an unwanted fall event, and sends alert when a fall happens.
Various solutions are proposed for solving this problem, each using different types of sensors. In this study, an artificial neural network is used and trained with videos of daily activities and samples of fall events. The used artificial neural network is a 3D Convolutional neural network, which is kind of a deep neural network. This network receives video-patches and extracts both spatial and temporal features of the video samples. Two datasets are used, named UR Fall Detection and SDU-Fall. The final accuracy of the system on the SDU-Fall dataset is 97.2%, which outperforms the state-of-the art studies on this dataset.
