This is the second part of my "SecuCam" project for the Project14's Digital Fever contest.
In the first part we connected a RaspiCam v1.3 to an Arduino MKR Vidor 4000, and did some basic image processing:
The camera image is taken from a RaspiCam v1.3. A Project14 logo along with an internet synced timestamp are applied as overlays. The resulting image was outputted to the Vidor 4000's MicroHDMI port.
Today we will use the TUL PYNQ-Z2 to implement Hardware Accelerated Motion Detection:
1. TUL PYNQ-Z2 and the PYNQ Framework
The TUL PYNQ-Z2 is a development board based on the Xilinx Zynq XC7Z020 FPGA SoC:
It features a 512 MB DDR3 RAM, a HDMI Input and a HDMI Output port and other interesting features.
The PYNQ-Z2 is also one of the "official" demo boards for the PYNQ (Python Productivity for Zynq) framework.
PYNQ (Python Productivity for Zynq) is a open-source framework from Xilinx that aims to make FPGA development easier.
PYNQ uses pre-built FPGA designs along with Python libraries, to allow exploiting the power of PL acceleration much easier.
Getting started with PYNQ on the is TUL PYNQ-Z2 easy. What we need to do is:
- download and flash a PYNQ SD card image on a microSD card
- connect a MicroUSB cable for power, and Ethernet cable for networking
- (optional) connect a HDMI Input (source) and Output (display)
After we power up the board, a Jupyter Notebook server will be started. We can access it on http://pynq:9090/:
Note: I have the board from the Road Test I did back in 2019. Feel free to check it out:
2. Hardware Accelerated Motion Detection
To do image processing with PYNQ will use OpenCV. For HDMI Input and Output we can use the Base overlay.
The PYNQ-ComputerVision overlays provide hardware accelerated versions of common OpenCV functionality:
- we can install the Computer Vision overlays by executing the following commands on a SSH session:
$ sudo -H pip3.6 install –upgrade git+https://github.com/Xilinx/PYNQ-ComputerVision.git
$ sudo apt-get install libopencv-*
- after a restart there will a new pynqOpenCV folder in the Jupyter with examples
To implement hardware (partially) hardware accelerated motion detection, I used the following algorithm:
- two consecutive Input Frames are used as input
- Gaussian Blur is applied for noise filtering
- the Absolute Difference of the two frames is calculated
- a Binary Threshold algorithm is used to transform the gray-scale image into a black-and-white image
- two Dilation iterations are applied to fill up small gaps in the white areas
- a Find Contours algorithm is applied on the resulting image
- the resulting Contours represent the areas in the image where Motion was detected
The parts that can be hardware accelerated are:
- the Gaussian Blur calculation, which is implemented as a 2D Filter - using xv2.filter2D()
- the Dilation - using xv2.dilate()
(note other steps may also be hardware accelerated, but it may require building custom Pynq overlays)
The algorithm is mostly inspired from the Use Python, Zynq and OpenCV to Implement Computer Vision article from Adam Taylor.
I implemented the algorithm a Jupyter notebook, in a step-by-step manner. The notebook is available in the following Gist:
A pure software and a (partially) hardware accelerated version of the algorithm were implemented.
The performance of the two versions looks like follows:
- software: ~10 FPS
- hardware: ~16 FPS
(note: it is not too spectacular result, but given that 3 of the 5 steps are still implemented in software it can be considered ok)
Here is a demo video showcasing the project:
Hope You enjoyed this project!