IndRNN based long-term temporal recognition in the spatial and frequency domain

This paper targets the SHL recognition challenge, which focuses on the location-independent and user-independent activity recognition using smartphone sensors. To address this long-range temporal problem with periodic nature, we propose a new approach (team IndRNN), an Independently Recurrent Neural Network (IndRNN) based long-term temporal activity recognition with spatial and frequency domain features. The data is first segmented into one second sliding windows, then temporal and frequency domain features are extracted as short-term temporal features. A deep IndRNN model is used to predict the unknown test dataset location. Under the predicted location, a deep IndRNN model is further used to classify the 8 activities with best performed features. Finally, transfer learning and model fusion are used to improve the result under the user-independence case. The proposed method achieves 86.94% accuracy on the validation set at the predicted location.