MUDIFEI: Multidimensional Feature Extraction and Interaction Model for Human Transition Action Recognition Based on Sensor Data

Currently, the recognition of human transition actions based on sensor data faces numerous challenges, including difficulties in feature extraction due to the complexity of transition action data, the issue of handling incomplete data samples, and the lack of publicly available datasets. To address these challenges, this paper proposes a sensor data recognition model for human transition actions based on multi-dimensional feature extraction and interaction, named MUDIFEI. Firstly, MUDIFEI introduces a multi-dimensional feature extraction method, including a temporal feature extension module, a channel-spatial feature extraction module, and a convolution enhanced feature extraction module. The temporal feature extension module can simultaneously mine local and global temporal features, enhancing the model’s capability to extract temporal features. Additionally, a channel-spatial feature extraction module is proposed to enhance feature extraction in spatial and channel dimensions through dimensional shift and a channel-spatial attention mechanism. The method also incorporates a convolution enhanced feature extraction module to capture convolutional features from different subspaces. By integrating features across temporal, spatial, channel, and convolutional dimensions, the accuracy of transition action recognition is significantly improved. Secondly, MUDIFEI presents an Euclidean distance network with multi-dimensional feature interaction module to extract Euclidean distance feature vectors. The proposed multi-dimensional feature interaction module enhances the model’s sensitivity to changes in movement by exploring the interaction relationships of accelerometer data and gyroscope data in both channel and spatial dimensions, thereby improving the model’s accuracy in recognizing incomplete transition action data. Experimental results based on two self-built datasets and one public dataset demonstrate that the proposed model effectively extracts features of human transition actions. The accuracy of transition action recognition reached 92.96% on the UCI-HAPT dataset, 94.67% on the TSA dataset, and 85.56% on the non-ideal transition action dataset HTSA.