Shop OBEX P1 Docs P2 Docs Learn Events
Using the Propeller with Open Computer Vision on the RPi 2 with Python — Parallax Forums

Using the Propeller with Open Computer Vision on the RPi 2 with Python

ratronicratronic Posts: 1,451
edited 2015-06-25 04:50 in Propeller 1
This is the method I used to install OpenCV-2.4.9 . OpenCV-3 has been in beta forever but is finally released.

I will have to wait for someone alot brighter than me to post instructions on how to install OpenCV-3.0.0 .

Here is the example I went from - http://robertcastle.com/2014/02/installing-opencv-on-a-raspberry-pi/

The difference is he is installing OpenCV-2.4.8 and this will show how to install OpenCV-2.4.9 .

So first I make sure I am in the 'pi' directory. Then in the terminal window enter -
sudo apt-get update

sudo apt-get upgrade

sudo rpi-update

After that there are 2 sets of dependencies that must be installed in the order shown. First in the terminal enter -
sudo apt-get -y install build-essential cmake cmake-curses-gui pkg-config libpng12-0 libpng12-dev libpng++-dev libpng3 libpnglite-dev zlib1g-dbg zlib1g zlib1g-dev pngtools libtiff4-dev libtiff4 libtiffxx0c2 libtiff-tools libeigen3-dev
When that finishes enter -
sudo apt-get -y install libjpeg8 libjpeg8-dev libjpeg8-dbg libjpeg-progs ffmpeg libavcodec-dev libavcodec53 libavformat53 libavformat-dev libgstreamer0.10-0-dbg libgstreamer0.10-0 libgstreamer0.10-dev libxine1-ffmpeg libxine-dev libxine1-bin libunicap2 libunicap2-dev swig libv4l-0 libv4l-dev python-numpy libpython2.6 python-dev python2.6-dev libgtk2.0-dev

You will see that some of that is already installed with the latest Raspbian but I installed them anyway as shown above. Then to get OpenCV-2.4.9.zip into your pi directory enter -
wget http://sourceforge.net/projects/opencvlibrary/files/opencv-unix/2.4.9/opencv-2.4.9.zip

When that finishes enter -
unzip opencv-2.4.9.zip

cd opencv-2.4.9

mkdir release

cd release

ccmake ../

When the Window pops up press c to configure and when it finishes you will have 3 pages of configuration.

If you know what to do then have at it. If you don't then I have posted my setup in the 3 pictures of my pages.

Use the arrow up and down keys to move to the next option. The 'enter' key toggles off/on and allows entry

to type information in the options that take it. You only needed to put 'Release' in the CMAKE_BUILD_TYPE

on page 1. Other than that you just need to toggle the on or off options so they match what is in my pictures.

When you have all of the options set press 'c' to configure. When that finishes press 'g' to generate.

When that is done then in the terminal window enter -
make -j4

The -j4 uses all of the cores for the build. The good news is that on a Pi 2 the build only takes around a hour and a half instead of several hours on a Pi 1.

When the above finishes enter -
sudo make install

At this time OpenCV will work with USB webcams but to make it see the RPi camera module

we need to setup the uv4l-raspicam driver including the uv4l-raspicam-extras which you can find here -

http://www.linux-projects.org/modules/sections/index.php?op=viewarticle&artid=14

After installing exactly as shown on that page, in a terminal window enter -
sudo nano /etc/uv4l/uv4l-raspicam.conf

Editing that file remove the '#' in front of width = 640 and height = 480. Also remove the '#' in front of nopreview = yes.

Then save the file ctrl-x, y, then hit enter. In the terminal window enter -
sudo service uv4l_raspicam restart

You are done! The raspicam driver using the restart method shown above you can edit the raspicam.conf file to enable camera effects

or different camera settings. I will soon post a Python program that uses trackbars to adjust some of the settings in your program.

I will post both a Python program for the Pi and a C program for the Propeller to have a servo pan/tilt camera track an object.

If anyone gets OpenCV installed I would like to know. Mainly I have been using it to detect and track colored objects and then

send the info via serial to the Propeller after each processed frame.
1024 x 561 - 88K
1024 x 561 - 99K
1024 x 561 - 45K

Comments

  • ratronicratronic Posts: 1,451
    edited 2015-06-15 19:11
    This Python program is initially set to detect and track red objects and send the detected object's

    x, y screen location (1 byte each) and the object size (2 bytes) out the Pi's serial port. Also the

    packet is preceeded by a 2 byte flag. This program first takes a full size camera image 640x480 and

    then puts the trackbars on that image. The program changes from a BGR to a HSV color space and allows

    you to adjust the upper and lower threshold for Hue Saturation and Value. In a day or two I hope to post a C program

    for my Activitybot that has a servo pan/tilt camera tracking the object that has been tuned in using this

    program. The frame size is set at 200x140 and the FPS looks to be over 20 bouncing up to 29.
    # detect red initially - adjustable upper and lower Hue, Saturation, and Value threshholds
    import numpy as np
    import cv2
    import serial
    import time
    
    def nothing(x):
        pass
    
    cap = cv2.VideoCapture(0)
    ser = serial.Serial('/dev/ttyAMA0', 115200)
    cv2.namedWindow('Frm')
    cv2.namedWindow('Thr')
    cv2.namedWindow('Th')
    cv2.moveWindow('Frm', 10, 100)
    cv2.moveWindow('Thr', 350, 100)
    cv2.moveWindow('Th', 700, 100)
    cv2.createTrackbar('HL', 'Th', 0, 179, nothing)
    cv2.createTrackbar('HH', 'Th', 3, 179, nothing)
    cv2.createTrackbar('SL', 'Th', 40, 255, nothing)
    cv2.createTrackbar('SH', 'Th', 255, 255, nothing)
    cv2.createTrackbar('VL', 'Th', 40, 255, nothing)
    cv2.createTrackbar('VH', 'Th', 255, 255, nothing)
    tt = str(0)
    cf = cv2.getTickFrequency()
    font = cv2.FONT_HERSHEY_SIMPLEX
    time.sleep(2)
    _, th = cap.read()
    cv2.imshow('Th', th)
    
    while(True):
    
        st = cv2.getTickCount()
        hl = cv2.getTrackbarPos('HL', 'Th')
        hh = cv2.getTrackbarPos('HH', 'Th')
        sl = cv2.getTrackbarPos('SL', 'Th')
        sh = cv2.getTrackbarPos('SH', 'Th')
        vl = cv2.getTrackbarPos('VL', 'Th')
        vh = cv2.getTrackbarPos('VH', 'Th')
    
        _, frame = cap.read()
        frame = cv2.resize(frame, (200, 140))
        orig = frame.copy()
        frame = cv2.blur(frame, (5, 5))
        hsv = cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)
        thresh = cv2.inRange(hsv, np.array((hl, sl, vl)), np.array((hh, sh, vh)))
        thresh2 = thresh.copy()
        contours, hierarchy = cv2.findContours(thresh, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
        max_area = 0
        
        for cnt in contours:
            area = cv2.contourArea(cnt)
            if area > max_area:
                max_area = area
                best_cnt = cnt
    
        if max_area > 0:
            M = cv2.moments(best_cnt)
            cx, cy = int(M['m10'] / M['m00']), int(M['m01'] / M['m00'])
            cz = int(max_area)
            cv2.rectangle(orig, (cx - 5, cy - 5), (cx + 5, cy + 5), (255, 255, 255), 2)
            cv2.putText(orig, tt, (0, 25), font, 0.9, (0, 255, 255), 2)
            ser.write('!~')
            ser.write(chr(cx))
            ser.write(chr(cy))
            ser.write(chr(cz & 0xff))
            ser.write(chr(cz >> 8))
    
        # Show it, if key pressed is 'Esc', exit program
        cv2.imshow('Frm', orig)
        cv2.imshow('Thr', thresh2)
        if (cv2.waitKey(1) & 0xff) == 27:
            break
        tt = str('FPS %d'%(cf / (cv2.getTickCount() - st)))
    
    # Clean up everything before leaving
    cap.release()
    cv2.destroyAllWindows()
    ser.close()
    
  • Bill HenningBill Henning Posts: 6,445
    edited 2015-06-16 07:41
    Dave,

    Looks very interesting!

    I'll be keeping a close eye on this thread.
  • TonyWaiteTonyWaite Posts: 219
    edited 2015-06-16 10:06
    Hi Ratronic,

    This is great - thank you for posting!
  • ratronicratronic Posts: 1,451
    edited 2015-06-17 11:57
    Here is one example of using the output from the above program with C on a Propeller

    Activity board (actually an Activitybot). As you can tell from the video I am posting that

    the light shining on a object makes a big difference on how well the filtered object shows up.

    So do not be afraid to play with the threshold values to match your lighting. I was going to

    post a video yesterday but I used a different room with different lighting and could not get

    the defaults set in the program above to pick up the object good enough. So I took the video

    in the room I made the program in. The LED on P27 toggles with each new received frame packet.

    https://youtu.be/CJMGuesRhqo
    /*   pan/tilt camera tracker   */
    #include "simpletools.h"
    #include "servo.h"
    #include "fdserial.h"
    
    void track(void);
    
    #define INW 10          /* inner window +/- pixel boundry */
    #define OUTW 40         /* outer window +/- pixel boundry */
    #define INUSEC 2        /* inner window boundry microseconds to +/- from servo setting */
    #define OUTUSEC 50      /* outer window boundry microseconds to +/- from servo setting */
    #define CENTER 1500
    #define MINSERVO 750
    #define MAXSERVO 2250
    #define PAN 8
    #define TILT 9
    fdserial *ser; 
    int area, x, y, z;
    int x1 = CENTER;
    int y1 = CENTER;
    
    int main() {
      set_directions(27, 26, 0b11);
      ser = fdserial_open(6, 7, 0, 115200);
      servo_set(PAN, CENTER);
      servo_set(TILT, CENTER);
      pause(500);
      fdserial_rxFlush(ser);
      
      while(1) {
        track();
        if (x < -OUTW) x1 -= OUTUSEC;
        if (x < -INW) x1 -= INUSEC;
        if (x > OUTW) x1 += OUTUSEC;
        if (x > INW) x1 += INUSEC;
        if (y < -OUTW) y1 -= OUTUSEC;
        if (y < -INW) y1 -=INUSEC; 
        if (y > OUTW) y1 += OUTUSEC;
        if (y > INW) y1 += INUSEC;
        if (x1 > MAXSERVO) x1 = MAXSERVO;
        if (x1 < MINSERVO) x1 = MINSERVO;
        if (y1 > MAXSERVO) y1 = MAXSERVO;
        if (y1 < MINSERVO) y1 = MINSERVO;
        servo_set(PAN, x1);
        servo_set(TILT, y1);
      }  
    }
    
    void track() {
      while(1){
        if (fdserial_rxCheck(ser) == 33) {
          if (fdserial_rxTime(ser, 1) == 126) {
            x = fdserial_rxTime(ser, 1);
            y = fdserial_rxTime(ser, 1);
            z = fdserial_rxTime(ser, 1);
            area = z + (fdserial_rxTime(ser, 1) * 256);
            x -= 100;  /* detected window frame size 200x140 */
            y -= 70;
            toggle(27);
            break;
          }        
        }      
      }      
    }
    
  • dgatelydgately Posts: 1,621
    edited 2015-06-17 19:04
    ratronic wrote: »

    If anyone gets OpenCV installed I would like to know. Mainly I have been using it to detect and track colored objects and then

    send the info via serial to the Propeller after each processed frame.

    Just got this working on my RaspberryPi B2+... I saw at least one compile failure, but continued with the installation (that takes a long time!). The detect red script runs (though I had to modify the raspiCam config file to rotate the image 180 degrees.

    It's detecting some cabinets behind me that are actually more brownish than the red screwdriver handle that I place in front of the camera. SO, I'll look at modifying the script next to see if I can get it to detect the screwdriver handle or some other red item. So far, so good!

    Thanks,
    dgately
  • ratronicratronic Posts: 1,451
    edited 2015-06-17 19:34
    Ah I knew somebody brighter than I am would try it out! Maybe you could figure out what is needed to install OpenCV-3.0.0. The last time I tried 3.0

    it would handle simple video stuff but would spit an error using findContours. You can loose the serial stuff to get a little better speed. The FPS

    printed on the video screen does not count for the delays that creep in between some frames by Raspbian. That installation of OpenCV-2.4.9

    will let you compile C++ programs too.

    Edit: The hue threshold determines the color with red being at 0. The value lower threshold helps with the amount of light in the room.
  • dgatelydgately Posts: 1,621
    edited 2015-06-18 09:40
    ratronic wrote: »
    Maybe you could figure out what is needed to install OpenCV-3.0.0. The last time I tried 3.0 it would handle simple video stuff but would spit an error using findContours.

    opencv 3.0 syntax for findCounters changed, slightly. It returns an image as well. I just updated 1 line of your example python code:
    Change the following line from:
        contours, hierarchy = cv2.findContours(thresh, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
    
    To:
        image, contours, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    

    With that, I don't get the errors with opencv 3.0!

    I think I figured out my detection of red, as well... I forgot that I'm using the NoIR Raspberry Pi cam, so everything is reddish. I'll switch back to the "normal" camera, and try again!

    dgately
  • ratronicratronic Posts: 1,451
    edited 2015-06-18 10:03
    This method of filtering the color you want is very simplified. But I was suprised that the RPi could handle vision processing ( with small windows).

    That was the error I was getting too many returned results but I had found an example on line using 3.0 that showed the same usage. Thanks for

    that dgately.

    Edit: Using this method of color filtering is not very good with simular colors in the same frame. You can also tighten the hue threshold range like
    set both upper and lower thresholds to 0 for a more pure red.
  • ratronicratronic Posts: 1,451
    edited 2015-06-18 14:37
    I just put back my saved image of a disk that had OpenCV-3.0.0 built on it and put your fix in the program. Thanks so much for

    pointing that out dgately 3.0.0 works here too. From now on I will be using it instead of 2.4.9. IIRC the setup for that has 7 pages.
  • ValeTValeT Posts: 308
    edited 2015-06-22 13:45
    ratronic wrote: »
    This method of filtering the color you want is very simplified. But I was suprised that the RPi could handle vision processing ( with small windows).

    Hi ratronic,

    I am using a Raspberry Pi Model 2 B for an Autonomous UAV competition at my school to identify images. Indeed, it is capable of complex image processing at a high resolution ( a 15.9 Megapixel photo takes about 1 minute to run through my program ). The most time consuming part is finding contours, especially since I am finding all contours in an object to create a mask as opposed to just the edge contours.

    Do you have any future plans for using OpenCV on the Raspberry Pi?
  • ratronicratronic Posts: 1,451
    edited 2015-06-22 14:04
    This is the first I have worked with the RPi and OpenCV so I am a beginner. I have an RPi 1 B+ attached to my Stingray robot and have a RPI 2 attached to an

    Activitybot. I have seen people that have RPi's and OpenCV to have thier robot read symbols to control it's next action. I am always looking for OpenCV things

    on the net. At the moment I am just researching to see what can be done.

    Edit: I use very small pictures to find contours for speed.
  • ValeTValeT Posts: 308
    edited 2015-06-23 04:58
    Cool. There is definitely a lot that OpenCV can do. I am currently expanding my program to use Machine Learning so that I can get better results during the competition. OpenCV has SVM support built-in ( as I understand it ) so you can do a lot of neat things with that. In addition, OpenCV has alphanumeric character recognition as well.

    And there are even more features coming in OpenCV 3.x ( for lack of a completely stable release, I am using 2.4.11 currently )!
  • ratronicratronic Posts: 1,451
    edited 2015-06-23 06:42
    ValeT I am now using OpenCV 3.0.0 it was finally released earlier this month I think. They are calling it OpenCV 3.0 Gold.
  • ValeTValeT Posts: 308
    edited 2015-06-24 12:47
    ratronic wrote: »
    ValeT I am now using OpenCV 3.0.0 it was finally released earlier this month I think. They are calling it OpenCV 3.0 Gold.

    Hahahahahaha. I really hope it is gold.

    Do you prefer it over 2.4.11? Is it stable? Does it run faster?
  • ratronicratronic Posts: 1,451
    edited 2015-06-24 13:04
    They made at least one change that made code written for the earlier 2.4.x series have to be changed to run on 3.0.

    I had some problems compiling C++ code with it. I had been using 2.4.9 for awhile and liked it. But I think I will try

    to do a new build of 3.0.0 and start new. Have any suggestion's to help me out?
  • ValeTValeT Posts: 308
    edited 2015-06-25 04:50
    ratronic wrote: »
    They made at least one change that made code written for the earlier 2.4.x series have to be changed to run on 3.0.

    I had some problems compiling C++ code with it. I had been using 2.4.9 for awhile and liked it. But I think I will try

    to do a new build of 3.0.0 and start new. Have any suggestion's to help me out?

    I wish I could help. I have had so much trouble with 3.0.0 and it close-to-ready betas, I'm just going to stick with 2.4.11 for now, sorry. But good luck and let us know how it turns out!
  • Sorry guys but this was hard to find and I need to bump it so I can find it again.

    FYI I have OpenCV working on my Pi B+, and the current goal is to make a balancing bot that can navigate through little mazes made with colored ribbon to find a blue box, hopefully in time for Maker Faire Detroit.

    Now off to change this avatar... wow
  • I thought you only need to do

    sudo apt-get install python-opencv

    to get opencv 2?  The python package pulls in everything else without fuss in my experience,
    in fact this is generally true for any complex library for which a python interface exists.
  • The version of OpenCV you get with using that method has a problem with reading from the RPi camera

    module in that OpenCV only sees 64x64 frames from the RPi camera module. USB webcams work with all of

    the versions of OpenCV. I would actually recommend using Marks method for a much easier and quick

    method of installing OpenCV. Just keep in mind that the frame size of the RPi camera module is 64x64

    for that version of OpencV.
  • No, Pi camera is the reason I use OpenCV, not seen any issue.  You have to set the
    size to 640x480:
    width = 640
    height = 480
    
    # Test with input from camera
    if __name__ == &#39;__main__&#39;:
    
    &nbsp;&nbsp;&nbsp; capture = cv2.VideoCapture (-1)
    &nbsp;&nbsp;&nbsp; capture.set (3, width)
    &nbsp;&nbsp;&nbsp; capture.set (4, height)
    &nbsp;&nbsp;&nbsp; print "setup"
    
    &nbsp;&nbsp;&nbsp; while capture.isOpened ():
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; print "grabbing"
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; capture.grab ()
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; ret, im = capture.retrieve ()
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; if ret:
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; if not track (im):
    &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; break
    &nbsp;&nbsp;&nbsp; print "exiting"
    
  • I just noticed that post's are not numbered anymore. In my first post near the end I show how to setup the

    RPi camera module to output 640x480 at boot up. But the earlier version of OpenCV-2.4.3 you get using

    the easy method to install has a problem with only the RPi camera module. It takes a later version to correct

    the problem of it only seeing 64x64 frames from the RPi camera module. Last night I finally compiled a C++

    program that is similar to the Python program I posted using OpenCV-3.0.0  (OpenCV 3.0 Gold). So I am

    going to continue exploring using the latest version and will post more information about it if anybody is

    interested. Keep in mind I am a noobie to this stuff.
Sign In or Register to comment.