Figure 2.

Overall model architecture of mEar. (A) The temporal mEar model processes raw triaxial accelerometer data (3 × 200 samples) to identify temporal gait events, i.e., ICs and FCs. It features a TCN architecture (tTCN) with residual blocks that have exponentially increasing dilation factors. Each residual block includes dilated convolutional layers, batch normalization, ReLU activation, and dropout layers. The TCN output is fed into an FCN with a sigmoid activation layer. (B) The identified IC events are used to extract accelerometer values for a full gait cycle, which are input for spatial-stride-characteristic regression using two spatial mEar networks. Both spatial models consist of a TCN architecture (sTCN), whose outputs are fed into an average pooling layer followed by a linear layer. Abbreviations: TCN: temporal convolutional network; FCN: fully connected layer; ReLU: rectified linear unit;; avg. pool.: average pooling; IC: initial contact; FC: final contact; SL: stride length; SW: stride width.