Thread Tools
This thread is privately moderated by Jack Crossfire, who may elect to delete unwanted replies.
Jan 04, 2015, 01:17 AM
Registered User
Jack Crossfire's Avatar
Thread OP

Path following without road markings

It was time to gather some test footage for a vision based path following algorithm. The only way was a complete setup with the actual camera in the actual position.

Glued a wide angle lens on it & mounted it to get the widest possible field of view.

Wrote a simple frame capture program & web app to preview the video. The Ralink 8192 is the only wifi dongle still in production, which supports an access point & it's terrible. It worked just enough to aim the camera.

The old battery went only 4 miles. Helas, hacking the ESC to directly drive the H bridge did not increase the range. It's now almost as short as the linear regulator.

Then, the raspberri pi's C library ended up being the 1st one in 10 years which can't write 64 bit files without extra C flags, so the recording ended after only 48 minutes. It still gathered a decent assortment of terrain.

The only section with a reasonable chance of machine vision working.

Most of it had very little difference between pavement & dirt.

Shadows would be a hard problem.

Following curbs was not realistic, because of the number of tree branches & parked cars on the road.

Path following can be a matter of differentiating between 2 materials. The side looking cam didn't cover enough area to differentiate materials in 1 frame, but it did cover as much area over several frames as a very high resolution cam from far away. Helas, differencing the time shifted video didn't reveal any difference in materials.

It didn't look good for side looking vision. After 8.4 miles of this, thoughts turned back to tracking the athlete, forward looking LIDAR, or forward looking vision with less precision.
Sign up now
to remove ads between posts

Quick Reply
Thread Tools