An Elderly Indoor Behavior Recognition Method Based on Improved SlowFast Network*
Junping He, Min Xiang, Zhao Xiaoxiang · Journal of Physics Conference Series · 2022
Abstract Due to the limited resources of home-based elderly care monitoring equipment, how to accurately identify the elderly behavior is a hot issue concerned by many equipment manufacturers. A behavior recognition method suitable for embedded equipment is proposed. This method first introduces the TPN module in the SlowFast network to capture the visual action tempo of the elderly and uses Faster-CNN to detect the elderly and objects in the video frames. Then in the AIA module, the accuracy of the elderly indoor behavior recognition is improved by strengthening the correlation between the elderly and the objects. Experiments show that the improved network can run stably at 18FPS on Jetson Xavier NX. The indoor behavior recognition accuracy for the elderly reaches 90.61%, which is 8.25% higher than the original SlowFast network.