Human activity recognition method combining improved TCN with Bi-GRU
Aiming at the problem of low recognition accuracy caused by incomplete feature extraction and insufficient gener-alization performance of traditional human action recognition methods,we propose an action recognition model based on deep learning methods.It improves the traditional temporal convolutional network(TCN),exponentially reduces the dila-tion rate layer by layer,and optimizes the residual structure of TCN,so as to extract the time between long-term interval da-ta in the shallow network and standardize network output.The reconstructed structure is further combined with the bidirec-tional gated recurrent unit(Bi-GRU)to extract local features of the data,and finally input to the fully connected layer to integrate the features and perform Softmax classification.Experiments show that the proposed model maintains a low number of parameters on the self-built dataset and the public dataset UCI-HAR,and the accuracy reaches 99.61％and 94.16％re-spectively,demonstrating reliable recognition performance.
human activity recognitioninertial sensortemporal convolutional networkbidirectional gated recurrent unit
万方数据
维普
