TCN-BiSRU-V2 fall detection model with performance evaluation and comparative analysis

With the increase in the global elderly population, fall detection has evolved as a critical domain of research aiming to ensure safety and autonomy for the elderly population. As existing fall detection models still face challenges in accurately identifying falls, particularly in complex scenarios or with diverse sensor data there is a significant scope for introducing better models for fall detection. In this work, we propose a novel hybrid deep learning-based model, the TCN-BiSRU-V2 model, which uniquely integrates Temporal Convolutional Network (TCN), Bidirectional Simple Recurrent Unit (BiSRU) and the GoogLeNet Inception V2 mechanism offering enhanced temporal modeling and computational efficiency for precise fall detection. The TCN component leverages temporal dependencies within the data to detect falls accurately, focusing on capturing the intricate temporal patterns inherent in human movements. BiSRU enhances computational efficiency by simplifying state updates and processing input sequences bidirectionally, effectively addressing the vanishing gradient problem and speeding up training compared to LSTM and GRU, which involve more complex gating mechanisms. Additionally, the model incorporates the Inception V2 mechanism to enhance data focussing, leveraging temporal dependencies, bidirectional processing and advanced features for accurate fall detection. The proposed model is evaluated on three benchmark datasets, namely SisFall, MobiFall and KFall, which consist of inertial sensor data achieving accuracies of 99.48%, 99.76%, and 99.85%, respectively. The outcomes show that TCN-BiSRU-V2 is superior to current techniques establishing new standards in fall detection. These findings not only underscore the efficacy of the model but also pave the way for its adoption in real-world applications, revolutionizing elderly care.