Human Keypoint-Guided Fall Detection: An Attention-Integrated GRU Approach

In recent years, the convergence of deep learning and medical advancements has seen rapid growth, reflecting an increasing societal aspiration for a superior quality of life. Falls, however, remain a principal threat causing grave injuries and mortalities, with the elderly being particularly vulnerable. This accentuates the imperative need for advanced computer vision based fall detection technologies. In this study, the maximum inter-frame difference method is first introduced to extract the keyframes from a video of the fall process. Subsequently, the skeletal keypoint detection algorithm via ViTPose++ is employed to derive comprehensive human skeletal keypoint coordinates. Then, the Gate Recurrent Unit (GRU) is introduced to recognize the inherent inter-frame correlation in fall detection. To further amalgamate spatio-temporal features, a fall detection algorithm via end-to-end neural network is proposed integrating the Attention mechanism with the GRU model. Ultimately, experiments on the public dataset for fall detection have affirmed that this enhanced GRU-Attention model secures an impressive accuracy rate of 99.28% with an AUC of 99.11% and improving the prediction efficiency by as much as 10 times.