Online Spatio-temporal Action Detection for Eldercare

Using AI technologies for assisted living of the elderly greatly facilitates care provision by caregivers and healthcare professionals. This paper proposes a lightweight model for real-time detection of human actions focusing on the elderly using a conventional RGB camera and an AI edge device. Our model analyzes and predicts the actions as the video frames arrive live from the camera and utilize spatio-temporal action detection to detect and locate the human actions with minimum latency. It can also apply to scenarios involving multiple people. We evaluate the proposed method using the popular public action detection dataset, AVA as well as an in-house self-collected dataset. The results show that our model can accurately detect the various actions in real-time at 15.2 fps using a resource-constrained edge device, offering significant potential for applications in various smart monitoring systems.