STAM-HAR:A Deep Learning Approach for Human Activity Recognition

Exoskeleton robots are a type of intelligent equipment that deeply integrates with human, and enhancing the recognition capabilities for human activities is crucial for the transition of exoskeleton technology from the laboratory setting to practical applications. For the case where the current attention mechanism does not effectively realize the human activities recognition(HAR), this paper proposes a spatio-temporal convolutional attention model named STAM-HAR. The combination of feature space convolution and temporal convolution networks constructs the architecture of STAM-HAR. The network structure is novel, compared with the traditional convolutional and recurrent neural networks, our model is able to fully explore the spatial information between individual sensors while extracting the time series information of the sensors more effectively. The efficiency and superiority of the model is demonstrated on the publicly available HuGaDB, a publicly available dataset, and on data collected by our team using the inertial motion capture device.