2.5. Post-Processing

During classification, each sliding window of 200 points with 20 points of separation was used to analyze the EMG signal, and then a vector with the probability of each gesture class was obtained, and only the most probable class was considered as the result of the classification stage. Then, the post-processing receives each of those class results, and a vector of labels is created by concatenating them. The vector of labels is finished when the number of sliding windows analyzed reaches the 5 s of recording. Then, we analyze the mode of every four labels, and the result is stored in a new vector of labels B∗, which is key to remove spurious labels that might appear during the classification results. In addition, we assign each those label results to a point in the time domain depending on the position of each sliding window. A sample of the vector of labels B∗ in the time domain is illustrated in Figure 6, where we can observe a set of noGesture labels, followed by a set of fist gesture labels, and again a set of noGesture labels. The ground truth A∗ can also be observed, which was obtained from the manual segmentation of the muscular activity that corresponds to a gesture. Finally, a recognition is considered successful if the vector of labels corresponds to the ground truth label, and if the vector of labels is aligned in time domain with the manual segmentation as illustrated in Figure 6. For this purpose, we used a minimum overlapping factor of ρ=0.25 as a threshold to decide if the recognition is correct. The overlapping factor is described in Equation (18),

where A∗ is the set of points where the muscle activity is located by the manual segmentation, and B∗ is the set of points where the gesture was detected by the model during post-processing.

Calculation of value ρ through overlapping among ground-truth and the vector of predictions. If overlapping factor for each EMG sample is more than ρ=0.25, then we consider that the recognition is correct.