Predictions of Movements by Online Analysis of Electroencephalogram with Dataflow Accelerators

Abstract :

Processing electroencephalographic data (EEG) for brain reading or brain computer interfaces is usually computational very expensive. This is very challenging, if one wants to develop mobile devices that can be used outside of the lab. In this presentation, we present our mobile processing unit that incorporates a Xilinx Zynq, which combines programmable logic and an ARM dual core CPU. Computational expensive tasks are executed on the FPGA by dataflow accelerators (e.g., filtering, spatial filtering and classification), while software-typical tasks are performed by the CPU (e.g., user interaction, data acquisition and slicing of the data). As a sample application, we show the prediction of voluntary movements of the right arm in 8 male subjects. We show that the classification performance does not differ from a pure CPU based solution, even though our FPGA based solution uses fixed-point arithmetic compared to floating point computations of a CPU. Furthermore, a considerable enhancement of processing speed could be achieved when comparing our system to an Intel® i-7 950 CPU. In future, we will use direct memory access in order to accelerate the processing even more, since data-transfer is the main bottleneck right now.