基于混合特征图卷积神经网络的人体行为识别方法OA北大核心CSTPCD

The existing methods for human activity recognition using wearable sensors could not capture the struc-tural information between the sampling points of time series effectively and might ignore the potential connections between samples.To address this issue,a graph convolutional neural network model with hybrid time-frequency and structural characteristics was proposed for human activity recognition.Firstly,the time-frequency characteris-tics of the original signal were obtained by wavelet-packet transform,and the spatio-temporal graph was further con-structed to extract the structural characteristics to identify the dynamic characteristics between the sampling points.The distance constraint was added to the structural characteristics to weaken the influence of long-distance neighbors on the central node on the spatio-temporal graph.Considering that the extraction of structural characteristics was greatly affected by the topological relationship of the spatio-temporal graph.The time-frequency characteristics of the samples to construct the input topology of the graph convolutional neural network were selected,and the time-frequency and structural characteristics were combined as the input features of the network.Finally,the input fea-tures propagated along the input topology to obtain the final classification result.To evaluate the performance of the proposed model,experiments were conducted on the WHARF and DataEgo datasets.Results in terms of F1 scores indicated that the proposed model outperformed existing convolutional neural network-based methods,achieving a maximum improvement of 19.58 percentage points on the WHARF dataset and 26.44 percentage points on the Da-taEgo dataset.It demonstrated that the proposed model could effectively enhance the capability of activity recogni-tion by exploiting dynamic characteristics.