Thread Tools
This thread is privately moderated by Jack Crossfire, who may elect to delete unwanted replies.
Oct 04, 2013, 01:14 AM
Registered User
Jack Crossfire's Avatar
Thread OP

Beyond PX4flow

Another bag of propellers & the PX4flow flew its last video.

Final PX4flow video (3 min 42 sec)

A few more configuration options added & the PX4flow was flying very consistently, but slowly. The horizontal velocity has to be limited to 0.25m/s. Vertical velocity has to be limited to 0.1m/s. The mane problem with the horizontal is what kind of carpet it has. The mane problem with altitude is it's clearly switching to the barometer more for high speeds & the barometer has much more lag.

The problem with the PX4flow is it's a dead end, for all the work that went into putting it into a complete flying system. Not having any definitive requirements makes it an endless guessing game, but the PX4flow is definitely not the path.

The requirements are some resemblance to an ideal system that can measure position from its acceleration alone, with some unknown resemblance to the same speed & precision of a human arm. It has to fly relative to the universe, with no dependence on any surroundings. If such an ideal system existed, it would be a robotics revolution.

The problem with open hardware is no-one gives out the suppliers of their parts. You have a schematic, but no way to obtain the obscure lens it uses without buying a complete PX4flow module from the open hardware seller. You can't fix it if it breaks. You can't modify it. You can't find alternative lenses from the same supplier. It's a reverse engineering nightmare. That's probably why there was a lot more confidence in open hardware being profitable, while open software had no such black box in a hidden supply chain.

The next step would be a normal angle lens & bigger sensor with extremely high resolution, very large motion vector search radii, hardware motion tracking, 80 megabits of extremely fast RAM, which would allow it to cover a wide range of altitudes & speeds, over a wide range of materials.

Digikey shows:

MT9P031I12STM 2.2um 60fps 2592 x 1944

MT9T031C12STC 3.2um 12-93fps 2048 x 1536

MT9P401I12STC 2.2um 60fps 2592 x 1944

The options with low quantities & user solderable packages are pretty crusty. Machine vision hasn't had the swelling of interest that green tech & 3D printing have gotten. You can get the latest & greatest renewable energy widget in a breakout board, but forget about any image sensors newer than 8 years old.

Would say that optical flow will always require a minimum distance from the walls that increases with altitude, making it always a dead end. The ideal system wouldn't.

The memory alone would be a major BGA job. The 1990's era TSOP chips wouldn't be fast enough to get the full 2592x1944 frames in & out.

A really beefy single board computer & a high resolution webcam would avoid the custom hardware. Dual core Intel Atoms still seem to be the top end.

There is a raspberry pi cam which does uncompressed 1920x1080 30fps. It's just a really slow computer. There is hope it may one day support GPU computing, but it's a slow process, taking many years to implement.

Going back to the old single board computer blog post, most have become vaporware, but the single board computer from Korea has been updated. The formerly 1.7Ghz quad core is now a new chip that houses 4 * 1.6Ghz + 4 * 1.2Ghz cores. There's no uncompressed camera port on it.

The single board computers are now in a race for power monitoring, energy efficiency monitoring, smart home improvement. It has replaced the fevers for home theater PC's & web servers in altoids cans, but there's still no interest in cameras.
Last edited by Jack Crossfire; Oct 04, 2013 at 01:54 AM.
Sign up now
to remove ads between posts

Quick Reply
Thread Tools