Figure 1.

Figure depicting the overall pipeline of the micromovement recognition/bite detection part of the algorithm. From left to right, the windowed 3D accelerometer and gyroscope streams of length $w_l$ are transformed into the 116-dimensional feature vector ${\mathbf{f}}_i$. Next, using an array of ten one-versus-one SVM the feature vector ${\mathbf{f}}_i$ is transformed into the 10-dimensional SVM prediction score vector ${\mathbf{v}}_i$. By applying a voting scheme to the one-vs-one scores of the ten SVM (i.e., to the ${\mathbf{v}}_i$ vector) we obtain $y_{i,mm}$ which indicates the micromovement label that corresponds to the ith window. Processing of additional sensor windows leads to the creation of the meal’s SVM score matrix $\mathbf{V}$ and label vector ${\mathbf{y}}_{\mathit{mm}}$. Furthermore, by processing windows of $\mathbf{V}$ with length $w_l^{\prime }$, the RNN outputs the probability $p_i$ that the given window sequence is a food intake cycle. Variables ${\mathbf{h}}_{i,j}$ and ${\mathbf{c}}_{i,j}$ are used to represent the ith hidden output and cell state of the jth LSTM layer, respectively. The rightmost part of the figure illustrates the local maxima search in the meal’s prediction vector $\mathbf{p}$. Variables $b_1$, $b_2$ and $b_3$ represent three detected bites, while $p_t$ represents the prediction threshold.