A Fast and Low-Impact Embedded Orientation Correction Algorithm for Hand Gesture Recognition Armbands

Abstract

Hand gesture recognition is a prominent topic in the recent literature, with surface ElectroMyoGraphy (sEMG) recognized as a key method for wearable Human-Machine Interfaces (HMIs). However, sensor placement still significantly impacts systems performance. This study addresses sensor displacement by introducing a fast and low-impact orientation correction algorithm for sEMG-based HMI armbands. The algorithm includes a calibration phase to estimate armband orientation and real-time data correction, requiring only two distinct hand gestures in terms of sEMG activation. This ensures hardware and database independence and eliminates the need for model retraining, as data correction occurs prior to classification or prediction. The algorithm was implemented in a hand gesture HMI system featuring a custom seven-channel sEMG armband with an Artificial Neural Network (ANN) capable of recognizing nine gestures. Validation demonstrated its effectiveness, achieving 93.36% average prediction accuracy with arbitrary armband wearing orientation. The algorithm also has minimal impact on power consumption and latency, requiring just an additional 500 μW and introducing a latency increase of 408 μs. These results highlight the algorithm's efficacy, general applicability, and efficiency, presenting it as a promising solution to the electrode-shift issue in sEMG-based HMI applications.

Keywords: embedded algorithm; hand gesture recognition; human–machine interface; surface electromyography; wearable armband.

Figure 1

Armband device structure: the representation shows the 7 channels (CH 1–CH 7) and the wired (I²C) and wireless (BLE) communication protocols implemented. The overlapped profiles represent the ATC values distribution for the nine hand gestures recognized by the system: wrist extension (WE), wrist flexion (WF), radial deviation (RD), ulnar deviation (UD), hand grasp (HG), thumb–index finger pinch (P2), thumb–middle finger pinch (P3), open hand (OH), and idle state (IS). The ATC profiles are based on the data acquired on 20 subjects in our previous work [28].

Figure 2

A general user wearing our custom sEMG armband. Considering its three degrees of freedom (i.e., DOF₁—rotation around the forearm, DOF₂—lengthwise position, and DOF₃—reversed insertion), the proposed embedded calibration algorithm is able to compute both DOF₁ and DOF₃ information, making the wearing of the device easier than searching for the exact reference position. The forward and backward shift of the armband along the forearm (DOF₂) is instead fixed one-third along the length from the elbow to the wrist in order to sense the muscular signals where muscle bodies are more accessible.

Figure 3

Classifier performance with different artificial shift conditions of the input data. All the metrics suffer a relevant decrease in their values when the armband is rotated by about ±30° from the reference position (0°), significantly impacting the reliability of the classification outcomes.

Figure 4

Analysis of activation profiles of Ninapro publicly available databases [39,40]. Attributes of each database are given in Table 1. For each candidate gesture for the calibration routine, the figure shows the distribution of RMS profiles across the sensing channels. The plotted profiles in each radar chart were normalized with respect to their maximum value. The highlighted gestures are those identified as suitable for the calibration routine, while the black cross markers are the points to be used to correct the reversal wearing. The last radar chart is the distribution of ATC activation profiles of our database [28].

Figure 5

Accuracy of the system for different resolutions (i.e., number of interpolation points, N_interp, between adjacent channels) of the orientation estimation algorithm.

Figure 6

Software (SW) and firmware (FW) interconnected operations for the calibration and online classification working phases. The green blocks represent the new steps for estimating the orientation of the device and for correcting the sensed values before performing the classification. All the algorithm computations are embedded on the armband modules’ MCUs while the SW package runs the graphical user interface (GUI), helping the user to follow the indications for a straightforward experience. The bottom timeline also reports the developer-defined calibration protocol timings and the (*) system latency measured during experimental validation.

Figure 7

Reversal assessment: the reversal condition is determined depending on the position of the UD peak w.r.t. the WE peak. In particular, the Mean Absolute Error (MAE) has been computed by shifting the UD profile forward and backward one channel and comparing it with the WE profile. If the error in the backward position (MAE_bw) is lower than the forward one (MAE_fw), the armband is worn in the correct non-reversed condition (case A), otherwise a reversal condition is detected (case B).

Figure 8

Shift assessment: the estimation of the armband shift only requires the processing of the WE profile. The acquired ATC values (1) are extended with 3 more points at its extremities (1a) to be interpolated with a cubic spline function (2). Then they are cross-correlated with the reference WE profile (3) to find the angular shift of the armband.

Figure 9

ATC signal correction during real-time operations. First, if Reversal = true, values are symmetrically swapped w.r.t. CH 1 (i.e., CH 2 with CH 7, etc.). Then, the signal undergoes steps 1a and 2 of the shift assessment process. Lastly, the signal is shifted the saved amount of degrees, and resampled with the original channel distance to obtain the seven values (IN 1–IN 7) needed as input by the ANN (4).

Figure 10

Validation protocol: (A) the armband is worn on the right forearm with arbitrary orientation; (B) calibration gestures are performed to determine reversal and shift conditions; (C) standard gestures are executed to verify the system performances. The testing protocol maintains the same organization of our previous study [28].

Figure 11

Representation of all the different orientations casually resulting from the test phase, including 23 in the standard wearing condition and 19 in the reverse wearing condition.

Figure 12

Classification performance of the system after the implementation in real time of the automatic orientation correction algorithm.

Figure 13

Classification performance improvements if the two pinch classes are merged and a moving window is used on the output class.