Motion Sensor-Based Fall Detection Using 1D-FCN

Falls are a leading cause of unintentional injuries, particularly among the elderly, often leading to severe health complications. Effective fall detection systems are essential to mitigate these risks. Fall detection systems that rely on vision and multimodal approaches face limitations such as lighting conditions, privacy concerns, and high computational demands. This paper introduces a sensor-based fall detection model using inertial measurement units (IMUs) data, specifically gyroscopes and acceleromenters from wearable devices, thereby offering benefits in versatility, privacy, and real-time deployment. This research proposes a novel fall detection approach utilizing a fully convolutional network (FCN) architecture optimized for time series data. The FCN model effectively learns features from raw sensor data, providing high accuracy and computational efficiency. The proposed method was evaluated on the UP-Fall dataset, including a comparison with a baseline LSTM model. Results demonstrate that the proposed FCN model surpasses existing methods, achieving 99.52% accuracy, 98.7% sensitivity, 99.65% specificity, 98.14% precision, and 98.38% F1-score, with significantly lower inference time and memory consumption. These findings underscore the potential of the proposed approach for practical, real-time fall detection applications, ensuring prompt intervention and reducing the adverse effects of falls for the elderly.