Gas measurement in Paralelni Polis

Why

Paralelni Polis has 3 hungry gas boilers and we want to keep an eye on amount of gas they eat. Doing gas reading in 2016 with pen and paper is more than cumbersome.

Getting things ready

sudo apt-get install ipython python-opencv python-scipy python-numpy python-pygame python-setuptools python-pip python-sklearn python-cssselect python-lxml
sudo pip install https://github.com/sightmachine/SimpleCV/zipball/develop
sudo pip install svgwrite
git clone git@github.com:ParalelniPolis/ppplyn.git
cd ppplyn

Test image recognition on images included in repository

./test.py
./images/input/camera_1449323843.png	1474.652
./images/input/camera_1449323833.png	1474.648
./images/input/camera_1449323854.png	1474.657
./images/input/camera_1449323901.png	1474.678
./images/input/camera_1449323870.png	1474.664
./images/input/camera_1449323822.png	1474.643
./images/input/camera_1449323906.png	1474.680
./images/input/camera_1450650840.png	1947.815
./images/input/camera_1449323880.png	1474.669
./images/input/camera_1449323838.png	1474.650
./images/input/camera_1449323864.png	1474.662
./images/input/camera_1449323917.png	1474.685
./images/input/camera_1449323912.png	1474.683
./images/input/camera_1449323896.png	1474.675
./images/input/camera_1449323875.png	1474.665
./images/input/camera_1449323828.png	1474.645
./images/input/camera_1449323848.png	1474.655
./images/input/camera_1449323859.png	1474.635
./images/input/camera_1449323891.png	1474.673
./images/input/camera_1449323885.png	1474.671

Tools for camera settings

v4l2-ctl --set-ctrl brightness=100
# This needs to be done after reboot
v4l2-ctl --list-ctrls
# List options for current webcam
v4l2-ctl --info

Detecting digits

Source image

This is raw image taken by the camera. Camera is positioned slightly above the meter to prevent any reflections.

Marker detection

We have 4 oragne markers in each corner of the meter. This color can be easily detected on the image.

Gas meter with fixed perspective

Knowing position of each corner, we can transform image to rectangle. This helps us estimate position of the digits we are looking for.

Digits detection

Now we can assume in which area of the image is each digit. These coordinates are hardcoded in the detector.

Single digits

Now we are ready find blobs (digits) in each rectangle. With some black magic we can ignore blobs which are not digits, reflections and crap we are not interested in.

Machine learning

Currently we have two classificators used for digit recognition.

• SVCDigitDetector - Detector based on LinearSVC from Scikit
• TemplateDigitDetector - Dumb detector substracting two images and measuring the difference

Feel free to implement any other detector, it just needs to have method detect_digit()

At some point I want to implement ideas mentioned in this article

Training dataset

This dataset is used for training of image recognition algoritm.

for i in {0..9}; do echo -n "Digit $i "; echo "`ls $i/*.png | wc -l` samples"; done
Digit 0      481 samples
Digit 1      452 samples
Digit 2       99 samples
Digit 3       79 samples
Digit 4      409 samples
Digit 5      142 samples
Digit 6      102 samples
Digit 7      294 samples
Digit 8      153 samples
Digit 9      420 samples

Testing dataset

Testing dataset contains images which are not prsent in the Training dataset

for i in {0..9}; do echo -n "Digit $i "; echo "`ls $i/*.png | wc -l` samples"; done
Digit 0       68 samples
Digit 1       77 samples
Digit 2       31 samples
Digit 3       65 samples
Digit 4      100 samples
Digit 5       94 samples
Digit 6       56 samples
Digit 7       91 samples
Digit 8       70 samples
Digit 9       99 samples

Testing digit detectors

Currently we have two ways how to detect digits. SVCDigitDetector (LinearSVC - best) and TemplateDigitDetector (template substraction - poor).

There is a simple tool which tests detectors with the testing dataset.

./test_detectors.py
svc_digit:0 template_digit:0
svc_digit:0 template_digit:0
svc_digit:0 template_digit:0
svc_digit:0 template_digit:0
svc_digit:0 template_digit:0
svc_digit:0 template_digit:0
svc_digit:0 template_digit:0
...
...
svc_digit:9 template_digit:0
svc_digit:9 template_digit:0
svc_digit:9 template_digit:8
svc_digit:9 template_digit:0
svc_digit:9 template_digit:0
svc_digit:9 template_digit:0
svc_digit:9 template_digit:8
svc_digit:9 template_digit:0
svc_digit:9 template_digit:8
SVCDigitDetector    99.3342210386%
TemplateDigitDetector   76.0319573901%

Physical setup

• ODROID-U3
• Microsoft LifeCam HD-3000
• 3D printed cover from MakerLab TBC
• LED strips
• Tons of epoxy glue