Defense against UAV's

"Keith W" wrote in message
...

wrote in message
oups.com...
[ SNIP ]
A UAV with realtime video image recognition and IR sensors is unlikely
to be especially cheap
Realtime video image recognition needs a source of video (probably a
wide-angle search camera + narrow angle scope with some decent
magnification for examining the suspicios contacts),

Problem 1 ) You have to process them to decide if they are suspicious

a decent CPU to do
the number crunching and a software to do the analysis. The first two
items are not particularly expensive. The software might take real
pains to develop, but afterwards the copies are free.

Understatement of the year
[ SNIP ]

If he can show the image processing and recognition problem to be easy, his
soon-to-be net worth will be more than that of Bill Gates. In fact, he'll
hire Billy just to supervise the programming staff to write the queuing
software for his executive bathroom.

Despite humongous amounts of research being done over many decades, general
computer vision remains an intractable problem. To illustrate, it may be
impossible for a vertical photograph (satellite) to differentiate between a
parking lot surrounded by a board fence and with a few cars parked on it,
and a large building with a flat roof and large roof vents, both with roads
nearby and under conditions of shadowing.

Now, in this case we'd certainly have a rules base codifying the knowledge.
But even restricting the problem to that of finding ships in the open ocean,
it's still not that simple. At a typical distance and altitude, a lot of
those ship lines are actually curves, so your algorithms need to recognize
smooth curves as part of a ship definition. Hmmmm, what else at sea is often
a fairly smooth curve?

I have a photo (8 1/2 by 11") of most of 4th MEB at sea, either 1990 or
1991. Fourteen vessels (LSTs, LPDs, LSDs, LPHs, one LHA, and a hospital
ship - no UNREP ships) are depicted. Going off the length of the LHA (shown
at a significant oblique), my estimate (very rough) is that the formation is
5-6 km across and perhaps 3 km deep. Even compressed like this - it's a
formation in time of hostilities that surely makes captains nervous - it's
still a collection with lots and lots of empty space. The colour contrast
and the wakes, the very calm conditions and excellent vizibility (light
haze) will at least allow a decent software to identify the ships as ships.
Leaving aside the hospital ship, I don't see that classifying most of the
vessels in the photograph would be anything other than a [very] difficult
recognition problem. A human can do it quickly, especially if cued with the
knowledge that everything is a USN gator, but it would be a pretty expensive
program that reliably typed each target.

One wonders too if the supposedly small and cheap UAV with the purportedly
inexpensive but sophisticated image recognition system is also fixing the
precise 3-D attitude of the airframe and hence the camera in order to allow
for estimating sizes of the objects in the picture, and _their_ attitude.
Forget relying on the horizon - in my picture you can barely make it out
because of haze. And it would have to be really precise data in order to get
good dimensional info.

What if you can't even see the wake, for one of several reasons? I'll give
the program three stars if it even correctly figures out what end of the
ship is which.

Now let's suppose that I am somewhat harsh in my analysis. Let's say that a
relatively coarse resolution picture and a basic analysis alerts the
software to "blobs of interest", and then the vehicle + camera is commanded
to do what it needs to do to get high-res images, and a better routine
analyzes these. Given some near optimal pictures - nearly side-on to the
vessel - you'd have something to work with. But in order to gauge size,
you'd need to be at some moderate altitude to have good geometry, under
which conditions superstructure begins to blend into the rest of the mass,
not be outlined against sky. In any case, with a large, detailed image of
the target, you now encounter other recognition problems e.g what details do
I ignore?

It is not a simple problem.

AHS

In article V%qfg.6218$JX1.2803@edtnps82,
says...
"Keith W" wrote in message
...

wrote in message
oups.com...
[ SNIP ]
A UAV with realtime video image recognition and IR sensors is unlikely
to be especially cheap
Realtime video image recognition needs a source of video (probably a
wide-angle search camera + narrow angle scope with some decent
magnification for examining the suspicios contacts),

Problem 1 ) You have to process them to decide if they are suspicious

a decent CPU to do
the number crunching and a software to do the analysis. The first two
items are not particularly expensive. The software might take real
pains to develop, but afterwards the copies are free.

Understatement of the year
[ SNIP ]

If he can show the image processing and recognition problem to be easy, his
soon-to-be net worth will be more than that of Bill Gates. In fact, he'll
hire Billy just to supervise the programming staff to write the queuing
software for his executive bathroom.

Despite humongous amounts of research being done over many decades, general
computer vision remains an intractable problem. To illustrate, it may be
impossible for a vertical photograph (satellite) to differentiate between a
parking lot surrounded by a board fence and with a few cars parked on it,
and a large building with a flat roof and large roof vents, both with roads
nearby and under conditions of shadowing.

Now, in this case we'd certainly have a rules base codifying the knowledge.
But even restricting the problem to that of finding ships in the open ocean,
it's still not that simple. At a typical distance and altitude, a lot of
those ship lines are actually curves, so your algorithms need to recognize
smooth curves as part of a ship definition. Hmmmm, what else at sea is often
a fairly smooth curve?

I have a photo (8 1/2 by 11") of most of 4th MEB at sea, either 1990 or
1991. Fourteen vessels (LSTs, LPDs, LSDs, LPHs, one LHA, and a hospital
ship - no UNREP ships) are depicted. Going off the length of the LHA (shown
at a significant oblique), my estimate (very rough) is that the formation is
5-6 km across and perhaps 3 km deep. Even compressed like this - it's a
formation in time of hostilities that surely makes captains nervous - it's
still a collection with lots and lots of empty space. The colour contrast
and the wakes, the very calm conditions and excellent vizibility (light
haze) will at least allow a decent software to identify the ships as ships.
Leaving aside the hospital ship, I don't see that classifying most of the
vessels in the photograph would be anything other than a [very] difficult
recognition problem. A human can do it quickly, especially if cued with the
knowledge that everything is a USN gator, but it would be a pretty expensive
program that reliably typed each target.

One wonders too if the supposedly small and cheap UAV with the purportedly
inexpensive but sophisticated image recognition system is also fixing the
precise 3-D attitude of the airframe and hence the camera in order to allow
for estimating sizes of the objects in the picture, and _their_ attitude.
Forget relying on the horizon - in my picture you can barely make it out
because of haze. And it would have to be really precise data in order to get
good dimensional info.

what's wrong with just sorting the targets in order of image size and
and allocating the UAVs on that basis. Do you think they will be
too worried if they get the hospital ship or an oiler instead of
the LPD?

What if you can't even see the wake, for one of several reasons? I'll give
the program three stars if it even correctly figures out what end of the
ship is which.

Does it need to know which end is which? I would just aim
for the center of mass.
Now let's suppose that I am somewhat harsh in my analysis. Let's say that a
relatively coarse resolution picture and a basic analysis alerts the
software to "blobs of interest", and then the vehicle + camera is commanded
to do what it needs to do to get high-res images, and a better routine
analyzes these. Given some near optimal pictures - nearly side-on to the
vessel - you'd have something to work with. But in order to gauge size,
you'd need to be at some moderate altitude to have good geometry, under
which conditions superstructure begins to blend into the rest of the mass,
not be outlined against sky. In any case, with a large, detailed image of
the target, you now encounter other recognition problems e.g what details do
I ignore?

It is not a simple problem.

You seem to be assuming that the enemy wants more information than
"3 big ships at coordinates x,y, speed X, on course B" For recon
information where you know the position of friendlies, that's probably
enougb to issue the targeting order.

Given traffic patterns in the Gulf, it ought to be pretty easy
to test whether your software can distinguish between 1000-foot
vessels and 100-foot vessels.

Mark Borgerson

"Mark Borgerson" mborgerson.at.comcast.net wrote in message
.net...
In article V%qfg.6218$JX1.2803@edtnps82,
says...


Given traffic patterns in the Gulf, it ought to be pretty easy
to test whether your software can distinguish between 1000-foot
vessels and 100-foot vessels.

Mark Borgerson



There are LOTS of 1000 ft vessels in the gulf, most are VLCC's

Keith



----== Posted via Newsfeeds.Com - Unlimited-Unrestricted-Secure Usenet News==----
http://www.newsfeeds.com The #1 Newsgroup Service in the World! 120,000+ Newsgroups
----= East and West-Coast Server Farms - Total Privacy via Encryption =----