Investigation on the Sampling Frequency and Channel Number for Force Myography Based Hand Gesture Recognition

Abstract

Force myography (FMG) is a method that uses pressure sensors to measure muscle contraction indirectly. Compared with the conventional approach utilizing myoelectric signals in hand gesture recognition, it is a valuable substitute. To achieve the aim of gesture recognition at minimum cost, it is necessary to study the minimum sampling frequency and the minimal number of channels. For purpose of investigating the effect of sampling frequency and the number of channels on the accuracy of gesture recognition, a hardware system that has 16 channels has been designed for capturing forearm FMG signals with a maximum sampling frequency of 1 kHz. Using this acquisition equipment, a force myography database containing 10 subjects’ data has been created. In this paper, gesture accuracies under different sampling frequencies and channel’s number are obtained. Under 1 kHz sampling rate and 16 channels, four of five tested classifiers reach an accuracy up to about 99%. Other experimental results indicate that: (1) the sampling frequency of the FMG signal can be as low as 5 Hz for the recognition of static movements; (2) the reduction of channel number has a large impact on the accuracy, and the suggested channel number for gesture recognition is eight; and (3) the distribution of the sensors on the forearm would affect the recognition accuracy, and it is possible to improve the accuracy via optimizing the sensor position.

A Review of EMG-, FMG-, and EIT-Based Biosensors and Relevant Human–Machine Interactivities and Biomedical Applications

“…The more channels, the higher the recognition accuracy. At the same time, they recommend using eight channels in the field of gesture recognition, which can achieve a satisfactory result [ 39 ].…”

Wearable super-resolution muscle–machine interfacing

“…Additionally, the number of sensor arrays depends on the number of ADC channels. There are two standard configurations: (1) one ADC for each sensor unit, which is simple but takes up a lot of space and is power-hungry with the high number of channels; (2) one ADC for multiple sensors by sharing a common reference or using a multiplexor for switching sensors, called a line scanning structure (Lei et al, 2021). Table 2 compares sensors in terms of size, high-density configuration, and non-invasive deep muscle detectabilitywhich are the three most important parameters of superresolution.…”

“…FMG signal acquisition can be achieved through different sensor designs, with some examples given in Figure 4 . The common ones are piezo-resistance or resistive polymer-thick-film-based (RPTF-based) sensor ( Dordević et al, 2011 ; Xiao and Menon, 2014 ; Radmand et al, 2016 ; Dwivedi et al, 2019 ; Liang et al, 2019a , b ; Barioul et al, 2020 ; Prakash et al, 2020 , 2021 ; Lei et al, 2021 ), capacitance-based sensor ( Meyer et al, 2006 ; Truong et al, 2018 ), piezoelectric-based sensor ( Han and Kim, 2013 ; Farooq and Sazonov, 2017 ), optical-fiber based sensor ( Fujiwara et al, 2018 ), wide range stretch sensor/conductive rubber ( Amft et al, 2006 ; Bifulco et al, 2017 ). RPTF-based FMG sensor accounts for more than half of the literature because this kind of sensor has a relatively simple read-out circuit: the core design includes a voltage divider with a buffer, which facilitates its application as a low-cost, high-density array configuration ( Xiao and Menon, 2019 ).…”

“…Because FMG senses the volumetric changes of muscle, a relatively low sampling rate is enough. Most researchers set the sampling frequency to 100 Hz, but a study by Lei et al (2021) shows that 5 Hz is enough for static finger movements. A low sampling rate might mean that FMG might have a limitation in classification latency.…”

“…Summarized from the literature, the trend behind MMI development is to have better signal qualities and features for applications, which requires allocating signals to muscles with higher accuracy. The high channel density of a sensor system has been proven to improve the overall accuracy in applications due to higher spatial resolution and the ability to differentiate proximal features ( Radmand et al, 2016 ; Grushko et al, 2020 ; Lei et al, 2021 ). However, the spatial resolution does not remain consistent in terms of depth because MMI sensing technologies have different capabilities to detect deep muscles ( Grushko et al, 2020 ; Nsugbe, 2021b ).…”