Thread Tools
This thread is privately moderated by Jack Crossfire, who may elect to delete unwanted replies.
Feb 08, 2015, 08:32 PM
Registered User
Jack Crossfire's Avatar
Thread OP
Discussion

How to test machine vision



After a month of commuting, a new electronics package was fabricated. The $5 Turnigy BEC was put to use. It was a piece of garbage that melted at 2A. The high energy BEC business is a mercenary business. They only contain a few dollars of parts, but digikey has a $25 minimum. Castle creations remanes the only supplier of a functioning high energy BEC, still charging $25 after 10 years.

A rising edge interrupt handler allows any pin on the STM32 to be used as an SPI bus, at reasonable bitrates. Piecing together forum dialogs revealed a way to access hardware SPI on the raspberry. There was a workable test program on http://www.raspberrypi.org/documenta.../spi/README.md by the name of spidev_test.c


There was never an organized way to stuff the truck.


The easiest way to test it was a yoga pad. That showed the steering feedback was in the right direction, with the right magnitude. It also showed the truck would automatically align parallel to the line instead of perpendicular to it. Line following robots did this for years, but it still feels like driving perpendicular to the line could happen, in some cases.


Managed to get the vision algorithm up to 10fps. Flaws in the vision algorithm immediately became known. The vision algorithm is a crude, horizontal edge detector with very large averaging area. A search for the strongest line is done in the region where the path is supposed to be.

The current standard practice for all machine vision is edge detection. Several gaussian blurs of different radius are applied. The difference between 2 gaussian blurs magically reveals the edges.

For all the guys who downloaded OpenCV, ran the chroma key demo on their raspberry pi, & forgot about it, getting useful results out of it probably requires taking a class in machine vision techniques. The most complete one is currently https://www.youtube.com/playlist?lis...H3AQjFKZS9XgZm

Helas, OpenCV faded from popularity before the boom in self driving cars, so path following was not a priority, only face detection.
Last edited by Jack Crossfire; Feb 08, 2015 at 08:42 PM.
Sign up now
to remove ads between posts


Quick Reply
Message:
Thread Tools