There is a handful of elderly abuse cases in Hong Kong’s elderly nursing homes.
However, the current possible solution of manually monitoring CCTV footage have some issues that make it unreliable: it wastes a lot of time and manpower; such a repetitive task may lead to human errors; in many real cases, the CCTV footage is not being monitored until elderly abuse incidents are being exposed; last but not least, those checking the footage in the elderly nursing home could be the abuser themselves.
Elderly abuse can result in both physical and psychological trauma, which is not a problem to be overlooked. Therefore, we invented Elder-AID, an artificial intelligence (AI) system used to identify suspected physical abuse towards elderly (like slapping or punching) through the given CCTV footage. It will then alert the superintendent of the organization of the elderly nursing home (instead of staff inside the center), who will double-check the snippets with suspected abuse and carry out corresponding measures to intervene as soon as possible.
Our physical abuse detection model makes use of the Long-term Recurrent Convolutional Network (LRCN) architecture. LRCN combines the use of Convolutional Neural Network (CNN) for image processing of each frame, and Long Short Term Memory (LSTM) to store data between frames of a video sequence. We used a dataset from Kaggle named the “Real Life Violence Situations Dataset”, and took 500 clips each from the Violence and Non-violence category as training data. We used a Google Colaboratory notebook, Python and related libraries like Tensorflow to make our model and prototype. After some evaluation, Elder-AID’s accuracy is 90.7%, which shows its ability to recognize physical abuse behavior.
Unlike some similar systems, Elder-AID focuses on detecting whether the action found in the footage contains violence, as physical abuse can come in many forms. Though other systems may have less false alarms, they will have more missed signals than our system, but our team didn’t want to take that risk of possibly missing on even one elderly abuse case.
Elder-AID works like an abuse scanner so it can be additionally added onto the existing elderly nursing home CCTV systems. For our current prototype and idea, the staff simply have to upload the CCTV footage onto an online platform, and our AI system will start the detection procedure, then send any alerts if necessary. The footage is kept secure to protect the elderly’s privacy. More importantly, with slight moderations, Elder-AID can also be expanded to be applied in a wide variety of environments where violence and abuse is likely to happen as well, for instance, prison, school campuses, elderly homes and streets. We also plan on optimizing the audio recognition feature to detect physical abuse through sounds like screaming and hitting in areas without CCTV supervision like toilets.
To conclude, we hope that Elder-AID can prevent elderly abuse from happening, and give them a peaceful life of retirement.