Microb Technology/2008/Embedded vision system

Current page: Microb_Technology / 2008 / Embedded vision system

Hardware

• Colibri PXA320 module + Orchid board from Toradex
• USB WebCam Logitech QuickCam Pro 5000

Software

• Windows CE
• Balor: Video Usb Class grabber (cross platform Windows/Linux)

(designed for Logitech QuickCam Pro 5000)

Balor Basic Features

• Read images in YUV/RGB format (MJPEG missing)
• Save captures in PPM format
• Save movies in MPEG1 (need ffmpeg)
• CLI-base interface to configure webcam parameters
• multi-windowing to display several pre-processing images
• very fast, should be real time

Image processing

• convert colors in TSV
• color filter on blue and red
• automatic color calibration
• use disc mask to filter balls
• build bounding box lists of matching shapes

Planning

TODO now

• dig deeper to explain some rejected balls
• temporal algorithms / ball tracking
• define protocol
  1. webcam is ready / calibration finished
  2. send robot position to webcam
• windows:
  1. image captures (convert MJPEG -> RGB)
  2. find how to stop webcam autoconfiguration

TODO later

• save images in JPEG, use FTP
• write a YUV -> RGB convesion ( using MMX ? )
• advanced algorithm exploration on shapes (hough, sobel, neuronal networks) to detect
  1. borders
  2. robot
  3. profiling

Colibri on Linux

• format flash and install uboot
• installation SDK / toolchain
• compilation of kernel for colibri
• find a minimal distribution
• add UVC module
• compilation and test of Balor
• validation with test scripts
• profiling and optimisation
  1. kernel configuration
  2. Intel Wireless MMX 2

Results

Source picture

Bounding box of detected balls are displayed. This is the worst case (a red ball is missing)

Color segmentation

Picture after blue/red/white filtering (thresholds got thanks to automatic calibration). This is the source image for shape matching algorithms.