Activity Recognition Using Dual-ConvLSTM Extracting Local and Global Features for SHL Recognition Challenge

For high precision estimation with SHL recognition challenge, we use a deep learning framework based on convolutional layers and LSTM recurrent units (ConvLSTM). We, UCLab(submission 2), propose the model combined two different ConvLSTMs. One ConvLSTM of convolution layers has large kernel size and the other has small kernel size. We expect that these two ConvLSTMs extract the different kind of features like global features and local features. Then we concatenate each ConvLSTM output, and input to fully connected layer. Finally, we convert output of fully connected layer to probability distribution by applying soft-max function. We finally determine that we input 10 axes of sensors to our model. The axes we use are 3 axes of linear acceleration, 3 axes of gyroscope, 3 axes of normalize gravitational acceleration and pressure of difference from previous value. As a result, using last 20% of lines for validation, predictions have around 0.931 of F1-score.