Deep Learning Based Surveillance System for Fall Detection in Elderly Population

Background: Falls are very dangerous to the elderly, and regular wearable or manual types of monitoring are intrusive and cannot be used in a continuous manner. The skeletal tracking with vision is a non-invasive approach, but the fall detection remains a challenge because of limited data sets, the class imbalance, and significantly different fall motions. Methods: The proposed Conv-BiLSTM model with Temporal Attention added to it is aimed at capturing spatial-temporal patterns in 30x99 skeletal sequences. The model is trained on URFD, GMDCSA24 and a Mixed Dataset with a normalization, Gaussian augmentation and class balanced learning. CNN layers are used to capture short-term motion signals, BiLSTM to capture long-term dependencies and attention to focus on important fall-related frames. Results: This model had a 91% test accuracy on URFD, 90% test accuracy on GMDCSA24 and a much higher test accuracy of 97% on the Mixed Dataset. The high accuracy of training (97–100) is consistently good and the better performance on mixed-dataset implies better generalization and less overfitting when diverse training data are used. Conclusions: The proposed system incorporates a privacy-friendly and accurate fall-detection system capable of running in a telehealth or elderly-care setting, smart-home safety system, and performs better than traditional CNN and LSTM baselines.