Capturing spatio-temporal patterns of falls individuals using efficient graph convolutional network model

Capturing spatio-temporal patterns of falls individuals using efficient graph convolutional network model

연구 분야: Databases

학회: Applied Intelligence

Falls are a major worldwide health concern among people, and the ability to detect and prevent falls can have significant implications for their safety and well-being. This paper uses an Efficient-Graph Convolutional Network (Efficient-GCN) model to extract discriminative features of fall actions. The proposed model is designed to handle the complex and dynamic nature of human movements during a fall event. The main problem in fall events is to capture spatiotemporal information that results from falls, plus the insufficient data size for training. To address this problem, we suggest a protocol to collect a fall dataset. The Kinect camera is used to collect skeleton data, which is then processed using the Efficient-Graph Convolutional Network (Efficient-GCN) algorithm to identify fall individual patterns. We present a comparative study between three methods Efficient-Graph Convolutional Network (Efficient-GCN), Support Vector machine (SVM), and k-nearest neighbor (KNN) for improving skeletal-based fall detection and deep convolutional neural network (DCNN) for depth data. To have a more global view we compare our results with public dataset on the three baselines variant noted as Baseline coefficient (Bx) where “x” denotes scaling coefficient, where Efficient-Graph Convolutional Network Baseline with coefficient 2 (Efficient-GCN-B2) on our collected dataset outperforms achieving 98,50% accuracy on the cross-subject. The Efficient-Graph Convolutional Network with coefficient 2 (Efficient-GCN-B2) algorithm achieves remarkably satisfactory results in detecting fall events on the robust representation which is a skeleton and Deep Convolutional Neural Network (DCNN) attains 97% on depth data.