Figure 7.

Diagrams visualizing the strategies for using a single codebook with multiple sensor channels as used by [19] (left) and us (right). With the left variant, subsequences from all channels are appended to a vector that is then used for codebook assignment, producing only one histogram feature vector per time-window and sensor. The variant on the right treats subsequences of all sensor channels equally, and assigns them using the same codebook, resulting in 3 independent histogram feature vectors.