Real-time Human Action Detection for Elderly Monitoring System

As the safety of elderly people living alone receives more and more attention, some works start focusing on the monitoring system for abnormal behaviors of the elderly. In this paper, we propose a simple and effective realtime human action detection method based on model fine-tuning for elderly monitoring system. At first, we built our own action dataset, where falling down is the most important abnormal behavior. We also utilize data augmentation technique to generate diverse training samples and prevent severe over-fitting. Then, we use our own action dataset to fine tune the improved ECO model. Through the part of the existing model parameters as the initial parameters of the new model, greatly accelerated the convergence rate of the model training. Finally, we apply the trained model to real-time human action detection. The experimental results prove that the proposed approach can accurately detect the abnormal behavior of falling down while ensuring the online real-time requirement.