Impact Detection in Fall Events: Leveraging Spatio-Temporal Graph Convolutional Networks and Recurrent Neural Networks Using 3D Skeletons Data

Traditional fall detection systems are highly effective at identifying that a fall has occurred, but they often struggle to pinpoint the exact moment of impact. This temporal precision is critical for distinguishing between actual falls and near-falls or other daily activities, which helps reduce false alarms and improves the efficiency of emergency response. This study addresses this gap by proposing a deep learning architecture designed to detect the specific moment of impact using 3D skeleton data.

The researchers treat human 3D skeleton data as a graph, where joints represent nodes and limbs represent edges. They employ a hybrid deep learning architecture consisting of three main components:

• STGCN (Spatio-Temporal Graph Convolutional Network): Used to capture both the spatial relationships between body joints and the temporal evolution of these movements.
• GRU (Gated Recurrent Unit): Integrated to process sequential data efficiently, allowing the model to determine the contribution of specific body joints to the fall event.
• BiLSTM (Bidirectional Long Short-Term Memory): Added to analyze fall sequences in both forward and backward directions, providing rich contextual information and handling variable-length sequences effectively.

The model was trained and evaluated on an improved version of the UP-Fall dataset, which provides 3D skeleton coordinates for various fall scenarios.

The proposed methodology achieves an accuracy exceeding 90% across diverse fall scenarios. By focusing on the specific moment of impact rather than the entire fall sequence, the model significantly reduces false positives caused by non-fall activities. The integration of GRU and BiLSTM layers enhances the model's ability to handle the temporal dynamics of falls, making it more robust for real-world applications where fall duration and speed can vary significantly.