Efficient DMA transfers management on embedded Linux PSoC for Deep-Learning gestures recognition: Using Dynamic Vision Sensor and NullHop one-layer CNN accelerator to play RoShamBo

Abstract

This demonstration shows a Dynamic Vision Sensor able to capture visual motion at a speed equivalent to a highspeed camera (20k fps). The collected visual information is presented as normalized histogram to a CNN accelerator hardware, called NullHop, that is able to process a pre-trained CNN to play Roshambo against a human. The CNN designed for this purpose consist of 5 convolutional layers and a fully connected layer. The latency for processing one histogram is 8ms. NullHop is deployed on the FPGA fabric of a PSoC from Xilinx, the Zynq 7100, which is based on a dual-core ARM computer and a Kintex-7 with 444K logic cells, integrated in the same chip. ARM computer is running Linux and a specific C++ controller is running the whole demo. This controller runs at user space in order to extract the maximum throughput thanks to an efficient use of the AXIStream, based of DMA transfers. This short delay needed to process one visual histogram, allows us to average several consecutive classification outputs. Therefore, it provides the best estimation of the symbol that the user presents to the visual sensor. This output is then mapped to present the winner symbol within the 60ms latency that the brain considers acceptable before thinking that there is a trick.