Resumen: A brain machine interface (BMI) is a system that enables communication and control of devices using brain signals. Technological advancements allow current embedded PCs to carry enough computational resources to process EEG signals and to develop embedded BMIs. In this work, the performance of the Odroid XU4 embedded PC is evaluated as a processing and control device for BMI, based on motor imagery (MI) paradigm, to obtain a portable, low cost, and trustworthy device. The Odroid XU4 is valued at $74 and has an Octa core processor (2GHz). In comparison, the Linux PC used has an Intel Core i7 processor (3.4GHz) and costs around $1500. To analyze the effectiveness on each device, a 2-class and 4-class motor imagery datasets were used (BCI Competition II-III). The first dataset comprises 280 trials sampled at 128Hz, filtered between 0.5-30Hz, C3, C4 and Cz channels. Instead, the 4-class dataset was sampled at 250Hz, filtered between 1-50Hz and comprises 60 channels; however, only 20 channels were used.

For the 2-class MI dataset, a Wavelet Transform (WT) was used as the feature extraction method. The resulting feature vector trained two classifiers: Support Vector Machine (SVM) and Multilayer Perceptron (MLP). For the 4-class MI dataset, the feature extraction methods used were WT and One-Versus-Rest Common Spatial Patterns (OVR-CSP). Training was computed using a Multiclass SVM classifier.

In the first, case, results show that both systems offer the same accuracy (92.8%) on each classifier. Nonetheless, the PC(t:0.009s) performs aproximately 6 times faster than the Odroi(t:0.06s). For the second dataset, WT showed 77% accuracy, while OVR-CSP achieved 89% accuracy. Processing times indicate the PC outperforms the embedded system approximately by 5 times using WT (PC:0.46s;Odroid:2.25s) and OVR-CSP (PC:0.47s;Odroid:2.64s). Despite noticeable differences, the Odroid system has proven its potential for the development of accessible fully embedded BMIs.

Autor(es):
Acuña, K.; Carranza, E. y Achanccaray, D.
Año: 2016
Título de la revista: Cybathlon Symposium
Ciudad: Zúrich, Suiza