Cheapest High speed camera up to 900 FPS!!

[HermannSW]

The highest framerate that can be achieved for capturing 640x128 stretched frames is 665fps!
Here is a sample frame, dcraw processed and converted to .png:


Half an hour ago I asked myself how fast 640x64 can be captured.
I was even able to capture with 1000fps(!), but interleaved with many frame skips.

So I did what this thread is about, capturing with 900fps, looks still coloful for 1000/900=1.1ms shutter time:


Now frame skip and delta distribution analysis are possible, I captured 7171 frames in 8s with Pi 2B and v1 camera:

Code: Select all

$ rm /dev/shm/out.*.raw
$ raspiraw -md 7 -t 8000 -ts tstamps.csv -hd0 hd0.raw -h 32 --vinc 1F --fps 900 -r "380A,0020;3802,78;3806,05DF" -sr 1 -o /dev/shm/out.%04d.raw 2>/dev/null
Using i2C device /dev/i2c-0
$ ls -l /dev/shm/out.*.raw | wc --lines
7171
$ 

Although 7171 frames got captured in 8s, /dev/shm is filled <50%. So you can capture 16s of 900fps video!

Code: Select all

$ df /dev/shm
Filesystem     1K-blocks   Used Available Use% Mounted on
tmpfs             448396 202888    245508  46% /dev/shm
$ 

Timestamp analysis can be done from tstamps.csv generated by above command line:

Code: Select all

$ cut -f1 -d, tstamps.csv | sort -n | uniq -c
      1 
      4 1097
      3 1098
      3 1099
   2825 1100
   4228 1101
      4 1102
      2 1103
      4 1104
      1 2200
     73 2201
     22 2202
      1 2203
$ 

So the frame delta time is 1101us ± 4us (908.2fps), with 1+73+22+1=97 frame skips.

Code: Select all

$ bc -ql
1000000/1101
908.26521344232515894641

This allows to find a longer range without skips:

Code: Select all

$ grep "^2" tstamps.csv 
2200,2,26521736979
...
2201,77,26521847039
2201,921,26522777046
2201,2801,26524847274
2201,4699,26526937313
2201,6484,26528902984
...
2201,6600,26529104394
$ 

There are no frame skips for 1898 frames in range 2801-4698.
This is how above 640x64 frame was generated from raw Bayer:

Code: Select all

$ cat hd0.raw /dev/shm/out.3000.raw >out.3000.raw
$ dcraw out.3000.raw 
$ double out.3000.ppm > out.3000.ppm.d
$ pnmtopng out.3000.ppm.d > out.3000.ppm.d.png
$ 

github repo contains lot of documentation, and tools to automatically create .ogg video and animated .gif from a given set of frames. You can see some animated .gifs created that way above in this thread.


P.S:
Even if taking the frame skips into account, the average framerate is still 896fps!

Code: Select all

$ echo "1000000/(((7171-97)*1101+97*2201)/7171)" | bc -ql
896.15421000650964618592
$ 

The timestamp file can be converted into a Matroska format. That allows to create a .mkv Matroska video from the .ogg video, where each frame gets exactly played at the correct timestamp. With such a .mkv video the frame skips can be neglected.

[HermannSW]

What I showed before was 640x32 video at 896fps average framerate with (5$) v1 camera.

Here you find details of 640x75 video at 1005fps average framerate with (25$) v2 camera
(more than double framesize, and above 1000fps, played at 30fps, 33.5 times slower than real):
viewtopic.php?f=43&t=109137&p=1307795#p1307795

[larapnr]

The highest framerate that can be achieved for capturing 640x128 stretched frames is 665fps!
Here is a sample frame, dcraw processed and converted to .png:


Half an hour ago I asked myself how fast 640x64 can be captured.
I was even able to capture with 1000fps(!), but interleaved with many frame skips.

So I did what this thread is about, capturing with 900fps, looks still coloful for 1000/900=1.1ms shutter time:


Now frame skip and delta distribution analysis are possible, I captured 7171 frames in 8s with Pi 2B and v1 camera:

Code: Select all

$ rm /dev/shm/out.*.raw
$ raspiraw -md 7 -t 8000 -ts tstamps.csv -hd0 hd0.raw -h 32 --vinc 1F --fps 900 -r "380A,0020;3802,78;3806,05DF" -sr 1 -o /dev/shm/out.%04d.raw 2>/dev/null
Using i2C device /dev/i2c-0
$ ls -l /dev/shm/out.*.raw | wc --lines
7171
$ 

Although 7171 frames got captured in 8s, /dev/shm is filled <50%. So you can capture 16s of 900fps video!

Code: Select all

$ df /dev/shm
Filesystem     1K-blocks   Used Available Use% Mounted on
tmpfs             448396 202888    245508  46% /dev/shm
$ 

Timestamp analysis can be done from tstamps.csv generated by above command line:

Code: Select all

$ cut -f1 -d, tstamps.csv | sort -n | uniq -c
      1 
      4 1097
      3 1098
      3 1099
   2825 1100
   4228 1101
      4 1102
      2 1103
      4 1104
      1 2200
     73 2201
     22 2202
      1 2203
$ 

So the frame delta time is 1101us ± 4us (908.2fps), with 1+73+22+1=97 frame skips.

Code: Select all

$ bc -ql
1000000/1101
908.26521344232515894641

This allows to find a longer range without skips:

Code: Select all

$ grep "^2" tstamps.csv 
2200,2,26521736979
...
2201,77,26521847039
2201,921,26522777046
2201,2801,26524847274
2201,4699,26526937313
2201,6484,26528902984
...
2201,6600,26529104394
$ 

There are no frame skips for 1898 frames in range 2801-4698.
This is how above 640x64 frame was generated from raw Bayer:

Code: Select all

$ cat hd0.raw /dev/shm/out.3000.raw >out.3000.raw
$ dcraw out.3000.raw 
$ double out.3000.ppm > out.3000.ppm.d
$ pnmtopng out.3000.ppm.d > out.3000.ppm.d.png
$ 

github repo contains lot of documentation, and tools to automatically create .ogg video and animated .gif from a given set of frames. You can see some animated .gifs created that way above in this thread.


P.S:
Even if taking the frame skips into account, the average framerate is still 896fps!

Code: Select all

$ echo "1000000/(((7171-97)*1101+97*2201)/7171)" | bc -ql
896.15421000650964618592
$ 

The timestamp file can be converted into a Matroska format. That allows to create a .mkv Matroska video from the .ogg video, where each frame gets exactly played at the correct timestamp. With such a .mkv video the frame skips can be neglected.

Hi Hermann,

i tried your codes because for a project in our firm I needed de maximum frame rate on the raspi cam possible.
But actually, the programm just gave me many failure messages and none of it worked .
I'm a beginner in programming, but I've looked everywhere and cannot find a solution to my problem.
Could you help me with this?
what do I need for the software, devices etc. to make this work?
I have the raspberry pi 3, camera modul v2, a SD Card with 32gb storage and a normal raspberry screen.
Thank you very much

[HermannSW]

Hi,

the program to use for framerates >200fps is raspiraw.
You find instructions on how to build and install in 1a here:
https://stamm-wilbrandt.de/en/Raspberry ... l#raspiraw

You need 6by9's dcraw for post processing the raw bayer frames:
https://github.com/6by9/dcraw

[karczas]

Has anyone tested if Raspiraw can scan just a single row of data with very high FPS? Theoretically, it should be able to do so around 50k FPS on Raspiraw 3B+ and OV5640 sensor. I tried to simply modify --height parameter in the sample scripts but it seems not to be working. I mean, I get the 640x1 image on the screen of my Rpi but it's pure green which I find weird.
I'm in the middle of going through the Raspiraw source coude to spot any limitations but maybe someone has already did it and I'm just wasting my time as it could not be done for whatever reason.

[HermannSW]

Hi,

> Has anyone tested if Raspiraw can scan just a single row of data with very high FPS?
>
not single row, but few rows.

> Theoretically, it should be able to do so around 50k FPS on Raspiraw 3B+ and OV5640 sensor.
>
Not really, v2 camera maxes out at 1007fps for 640xH modes, v1 camera at 750fps.
v2 camera maxes out at 998fps for 1920x39 mode:
https://www.raspberrypi.org/forums/view ... 2#p1412292

There seems to be some resource problem on CSI-2 bus responsible for that.

So you cannot get more than rougly 1000fps with v2 camera.
But you can get up to 100,000eps(!) frame (exposures per second) with v1 camera global external shutter.
https://github.com/Hermann-SW/Raspberry ... al_shutter

This is a 9000eps frame of an airgun pellet flying through two 5000lm leds (bottom+top):

[karczas]

Thank you HermannSW for clarification. Please, let us know when you have managed to catch the passing bullet photo with your technique.

[HermannSW]

A lot has happened since my last posting, you can find in the updated github repo:
https://github.com/Hermann-SW/Raspberry ... al_shutter

I even captured an airgun pellet separately, but only on small part of the 2MP frame. The right image is 1:1 taken from a 2MP frame captured (flying with 100m/s, frame had color issue at that time):


In the meantime I developed a technique of creating a kFPS video from kEPS frame, this is a sample 5000fps video of an airsoft pellet flying with 36m/s, played at 5fps, 1000× slower than real:


Fast (software) camera panning while creating the video from 2500eps (exposures per second) frame captured of airsoft pistol with one blue dot reveals that pellet did rotate fast (500rps or 30000rpm) after intentional collision with a lath:


Finally record framerate has been improved to 20000fps, I am just writing down the details, they will be on repo README.md today. A black propeller rotates fast over a black circular disk cardboard (invisible), only white marker at blade tip seems to rotate. The 7µs long flashes were 43µs apart, so flash frequency was 20KHz, and captured frame was a 20000eps frame. The generated 20000fps video is played at 20fps, 1000× slower than real (43 white markers are captured in circle):


P.S:
github repo is up to date now, contains description of new techniques and new tools for creation of 20000fps videos: