Research on abnormal behavior recognition of the elderly based on spatial-temporal feature fusion

Abstract

With the deepening of the aging society, the health care of the elderly and disabled people has been widely concerned by society. It is of great practical significance to monitor the physical conditions and abnormal behaviors of elderly and disabled people based on sensor devices to prevent diseases and dangerous behaviors. However, most of the current algorithms (classical machine learning algorithms, CNN, RNN, LSTM) are only limited to the traditional Euclidean space, with little expansion in Non-Euclidean space. The translation invariance limits the expression ability of spatial data. The signals collected by the wearable sensor based on human abnormal behavior recognition often come from different parts and different directions of the human body. The traditional Euclidean algorithm is used to solve the problem, which makes the classification results easy to ignore the spatial characteristics of the sensor signals. The emergence of graph neural networks provides a new method to solve this problem. A large number of studies have proved that the graph neural network has a great role in dealing with Non-Euclidean space problems. Therefore, this paper proposes a new method for abnormal behavior recognition of the elderly based on spatiotemporal feature fusion (SCGAT). The Non-Euclidean spatial features of sensor signals are extracted by graph attention neural network, and the classification model is constructed by combining the extracted time domain features of the time-series coding network. In the experimental results, the proposed method achieves better classification results than most of the current algorithms.