A study on constructing an elderly abuse detection system by convolutional neural networks

With population aging accelerating, people are currently facing a new challenge which is until now unexplored: elderly abuse. According to some statistics provided by World Health Organization, one in six elderlies, aged at least 60 years old, is a victim of physical offense by their relatives and caregivers. Elder people face many types of abuse. This work focuses on the physical abuse which is defined by the infliction of pain on a person. Physical abuse can severely damage a person, sometimes leading to long-term psychological consequences, hospitalization, and death. The contribution to solve this problem is as follows. A dataset is first built by collecting elder abuse videos. Second, the dataset are applied over three different networks: the standard 3D convolutional neural network (3D CNN), the 3D residual convolutional neural network (R3DCNN) and the factorized 3D convolutional neural network based on the residual network 'R(2 + 1)D CNN'. Lastly, this paper introduces a new preprocessing method called the repeated frames extraction that has been shown to be efficient for action recognition. The project has been concluded with satisfying results with accuracies of 99.21%, 84.37%, and 85% for training, validation, and testing, respectively, on the standard 3D convolutional neural network.