Tie-fighter's profile - activity

Hello everybody :)

I started developing a computervision application (stitching, object detection, background substractor) using a Raspberry Pi.

I profiled my code and it seems that a lot of time is wasted with copying the images.

Are there any well supported/documented alternatives to the Raspberry Pi that would provide better performance? (-> higher framerate and/or image size)

On a sidenote: How big is the performance impact of NEON support and multiple cores? Is there are roadmap for supporting CUDA on ARM? (https://en.wikipedia.org/wiki/Tegra_2#Upcoming_releases)

The Ouya seems to have a good cost-performance ratio. Opinions? https://en.wikipedia.org/wiki/Comparison_of_single-board_computers

You are correct. I am using Raspian.

I get the following output when compiling opencv:

[ 18%] Built target opencv_flann
[ 18%] Built target opencv_highgui_pch_dephelp
[ 18%] Built target pch_Generate_opencv_highgui
Linking CXX shared library ../../lib/libopencv_highgui.so
/usr/bin/ld: /usr/local/lib/libx264.a(common.o): relocation R_ARM_MOVW_ABS_NC against `a local symbol' can not be used when making a shared object; recompile with -fPIC
/usr/local/lib/libx264.a: could not read symbols: Bad value
collect2: ld returned 1 exit status
make[2]: *** [lib/libopencv_highgui.so.2.4.9] Error 1
make[1]: *** [modules/highgui/CMakeFiles/opencv_highgui.dir/all] Error 2
make: *** [all] Error 2

I've been following this guide. What am I missing? :/

I am developing for an embedded ARM system and I am looking for advice on improving overall performance.

Specifically:

What should one avoid when developing for performance? (e.g. what functions are slow)

How do you find out where you can improve your code? How to profile a opencv project under Linux?

What are fast algorithms for object detection?

Are there any code examples of fast implementations?

I would like to detect and track people using a Raspberry Pi, Model B v2 (512MB RAM) and a Logitech C310 webcam on a pan/tilt mount.

I experimented with the BackgroundSubtractorMOG2 which worked quite well, but stitching (to build a background for the entire range of vision) is to slow.

Now I am considering:

1. Using BackgroundSubtractorMOG2 without stitching and only moving the camera in discrete steps (-45° , 0°, +45°), treating the three FoVs seperately.
2. Using a fast algorithm for object detection. Maybe LBP?

Is it possible to detect people on a Raspberry Pi with LBP or some other algorithm in real time? Can a neural network be fast enough for this too?

I recompiled with ffmpeg but forgot to make install

-_-

This is solved, but I am not allowed to close the question yet... It didn't work so I recompiled but forgot to "make install" -_- Everything works fine now...

For ccp-examle-bgfg_gmg even the original mjpeg videos work... it seems to be a problem with compiling. I do not understand, wh...

I am running out of ideas... are there any fool-proof video samples I could try? Maybe it's a problem with the code.

To what codec/format?

I would like to use video datasets to test my program, but unfortunately it does not work.

How do I encode them for the current version of opencv?

PS: It seems to be a problem with cmake (OpenCV examples work but own project doesn't). I am using the tutorial CMakeList.txt that does not compile including the codecs. Can you give me any pointers on how to correctly do this?