Deep learning algorithms, such as CNN (Convolutional Neural Network), could provide high 8 accuracy for great number of applications including video analytics for surveillance and 9 automotive. Considering processing speed and energy efficiency, FPGA is a good hardware to construct customized CNN solution.

In this demo session, we want to benefit from hardware technology, and show a fast speed 12 and accurate video analytics system using state-of-the-art deep learning algorithms running 13 on low power FPGA. This system could process 16 channels of continuous input video 14 with the resolution of 1080p. Two functionalities could be easily switched by just clicking a 15 button in this live demo: one for vehicle, non-motorized vehicle, and pedestrian detection, 16 tracking, and attributes analytics; and the other for face detection and recognition. The deep 17 learning algorithms used are SSD and densebox for two kinds of objects’ detection, which 18 have state-of-the-art accuracy. The FPGA used is Xilinx MPSoC ZU9, and the whole 19 board including this FPGA only cost about 50 Watts with Peak performance at 5.6 TOPS.

There are 3 things that audience could enjoy with our demo. The first one is sending 16 21 channels of 1080p videos from laptop to our FPGA board by Ethernet. The FPGA board will 22 run deep learning algorithms and display the results on monitor. The second one is using 23 camera to capture the real-time video in the conference hall, using FPGA analyze the video, 24 and displaying the face recognition results on the screen. The third one is deploying any 25 CNN network belonging to our architecture specification range on our FPGA system within 26 several seconds. This is because we have design a customized deep learning processing unit 27 on FPGA which could accelerate most of the CNN models. Generally speaking, the audience 28 will find fast speed, low power, and high flexibility deep learning processing with our 29 hardware design on FPGA, especially in video analytics application.