Figure 2.

The generic model architecture of the deep learning model to predict initial contacts (ICs) and final contacts (FCs). The inputs are the accelerometer and gyroscope data from a single inertial measurement unit, which are fed to a temporal convolutional network (TCN) (left). The TCN consisted of repeating residual blocks (ResBlocks) with exponentially increasing dilation factor (middle). Each ResBlock was built from two sequences of a convolutional layer (Conv), batch normalization layer (BatchNorm), a rectified linear unit activation layer (ReLU), and a dropout layer (DropOut) (right).