Skip to content

Latest commit

 

History

History
103 lines (84 loc) · 5.24 KB

README.md

File metadata and controls

103 lines (84 loc) · 5.24 KB
Raw

SharkCV

FRC-focused Python implementation of OpenCV and NetworkTables designed for a coprocessor.

Linux Installation

To install all necessary dependencies run the following:

$ sudo apt-get install python-numpy python-opencv python-setuptools
$ sudo easy_install pip
$ sudo pip install pynetworktables

You may want to consider installing a mDNS client to interface with NetworkTables as well:

$ sudo apt-get install libnss-mdns

Usage

Command-Line Arguments

To see a full list of command-line arguments run:

$ python SharkCV.py -h

There are arguments for various input types, output types, video input settings, and webcam settings.

Wired Webcam Input - Basic

$ python SharkCV.py [module.py]
  • By default SharkCV will use the first webcam plugged in but you can specify it manually with an argument.
  • Root access is sometimes required for webcams at /dev/video*.

Wired Webcam Input - Advanced

Here is an example that sets some video and webcam options:

$ python SharkCV.py -vw 320 -vh 240 -wb 0 -wh 127.5 [module.py]
  • FPS improves dramatically when resolution is reduced. Setting webcam resolution with arguments (when supported by device) will yield faster processing than resizing with OpenCV.
  • FPS can be greatly affected by webcam exposure time (and brightness if it affects the exposure).
  • Webcams frequently do not obey settings from OpenCV/V4L (-w* arguments).

MJPG Input Stream

Thanks to Mike Anderson mjpg-streamer has been compiled for the roboRIO and it is possible to stream the same webcam to both Smart Dashboard and SharkCV. See below for MJPG output.

$ python SharkCV.py -im [url] [module.py]
  • Video and webcam options will be ignored here, mjpg-streamer input_uvc.so must configured separately.
  • Be mindful of the ports allowed by the FRC FMS when configuring mjpg-streamer.

File Output

SharkCV supports various output formats including images and videos.

$ python SharkCV.py -oi image.png [module.py]
$ python SharkCV.py -ov webcam.avi [module.py]
  • Output filenames are processed through datetime.strftime() so they support % date notation.
  • File output is expensive and will cut your FPS significantly, especially on devices with poor throughput like the Raspberry Pi.

MJPG Output Stream

$ python SharkCV.py -oj [module.py]
  • Browsing to your device's IP in a browser will serve an HTML page with the MJPG stream.
  • HTTP server port is configurable but be mindful of the ports allowed by the FRC FMS.

Module Construction

Here is an example methodology you can use to calibrate your SharkCV module/algorithm.

Input Setup

  1. Decide what kind of input you will be using: local webcam, IP webcam, or static images.
  2. Decide on what brightness/contrast/exposure/etc settings are best for your webcam. If tracking retro-reflective tape consider turning your brightness and exposure down for a darker image to give the tape more contrast.
  3. If your image is upside-down (due to webcam mounting) you can use rotate() to correct it.

Threshold Calibration

  1. Set up SharkCV to use your desired and have it output timestamped images. Example:
$ python SharkCV.py -oi "captures/%Y%m%d-%H%M%S-%f.png" [module.py]
  1. Use an image editing such as GIMP to color-drop the item you want to track. Keep track of the min/max HLS/HSV values. Subtract some amount off the minimum and add some to the maximum to give yourself some variance tolerance.
  2. Construct your module/algorithm using threshold() and contours_draw(). Using the same images you captured above run them through your module and have it output more timestamped images. Example:
$ python SharkCV.py -ii "captures/*.png" -oi "captures_processed/%Y%m%d-%H%M%S-%f.png" [module.py]
  1. Check that the contours drawn on the images are what you want.

Other Operations

  1. Consider using resize() before doing anything. SharkCV will process each frame much faster at lower resolutions. Power-of-two reductions seem to be the fastest.
  2. Consider using blur()/blur_gaussian()/blur_median() first to smooth out your image (if necessary).
  3. Consider using dilate()/erode() before finding contours. Warning, these operations can get expensive at large sizes or large number of iterations.
  4. Consider using contours_filter()/contours_sort() before doing any kind of expensive operation on all contours.

Credits

License

SharkCV is under GNU GPL v3 to ensure any derivatives are also open source.