Using Auto ML to classify how many parking spaces are there

This is an implementation of the AutoML tutorial for getting the count of parking spaces in front of the house I live in.

Collecting data

I used old Raspberry Pi with USB cam. Using Raspberry cam is also an option, but I had issues with the length of the cable. The script is executed via cron:

*/10  *	* * *	root	bash /home/pi/snap.sh

Where snap.sh contains following:

#!/bin/bash
mount | grep "/mnt/parking_photos" >/dev/null || mount /mnt/parking_photos
fswebcam -D 2 -S 20 -F 10 -r 1920x1080 --no-banner `date +"/mnt/parking_fotos/%y%m%d_%H%M%S.jpg"`

Folder parking_photos is a NFS shared volume. You do not need that once you are willing to save the photos on a local filesystem.

Info about parameters and arguments could be found at raspberrypi.org. In case your use case is the same I would suggest using at least -F 10. I was not interested in the things which moved in the picture and this helped me to partially blur them away.

Preprocessing

Removing dark pictures

Copy pictures captured by Raspberry to parking_photos_raw folder. Then execute filter_dark_raw_photos.sh to remove pictures which are just too dark. It will start a lot of shell scripts in parallel. Overall it will compress the picture to 1x1 pixel and check if brightness is higher than 20:

if [ `convert $1 -colorspace hsb -resize 1x1 txt:- | grep -o 'hsb\(.*,.*,.*\)' | grep -o '[0-9]*%)' | grep -o '[0-9]*'` -gt 20 ]; then
  echo $1
else
  rm "$1"
fi

Preparing testing data

This is the most tedious part. Just go trough the pictures and divide them into folders in data directory. Your data folder may differ, but I had 1, 2, 3, 4 and i, full. Each of them tells how many parking spaces are empty. i stands for disabled parking space.

Creating list of pictures

AutoML accepts the list in the format described in the documentation. Once you upload the pictures in to GCS bucket, you can use script data/generate_list.sh to get the list. It will divide the pictures into 80% of training, 10% for testing and 10% for validation.

Howto

The steps I took are the same as described in tutorial

Go to automl directory & use:

vision_classification_create_dataset.py to create the dataset,
import_images_to_dataset.py to import pictures,
vision_classification_create_model.py to create model,
list_model_evaluations.py to see if this all was worth it.

Do not forget to fill .env file in automl folder.

Results

Use python3 vision_classification_predict.py testing_photos/210705_161002.jpg to evaluate the testing pictures:

file_path = "testing_photos/210705_101001.jpg":

Prediction results:
Predicted class name: 2
Predicted class score: 0.44264715909957886
Predicted class name: 3
Predicted class score: 0.4311576187610626

file_path = "testing_photos/210705_140001.jpg":

Prediction results:
Predicted class name: 2
Predicted class score: 0.5498601794242859
Predicted class name: 3
Predicted class score: 0.4054759740829468

file_path = "testing_photos/210705_161002.jpg":

Prediction results:
Predicted class name: i
Predicted class score: 0.8651795387268066

As you can see, the score is not as high as desirable. Once you need to know if you can park the car, this number seems to be enough.

Camera holder

If you need to place the camera somewhere reasonable, you can use 3D printed holder. It's the holder folder.

Comparing with Tensorflow implementation

To have something to compare AutoML with, I tried standard Tensoflow tutorial on image classification. I had to change few constrains:

Large size of the batch did not help with better accuracy
Adding image shifting did not help either
Enlarging meta arguments like the count of Conv2D layers did help
Enlarging input image size helped a bit as well

Overall results were sufficient:

210705_101001.jpg image most likely belongs to 1 with a 98.84 percent confidence.
210705_140001.jpg image most likely belongs to 3 with a 82.20 percent confidence.
210705_145001.jpg image most likely belongs to i with a 80.34 percent confidence.
210705_153001.jpg image most likely belongs to i with a 100.00 percent confidence.
210705_154002.jpg image most likely belongs to i with a 99.99 percent confidence.
210705_155001.jpg image most likely belongs to i with a 99.72 percent confidence.
210705_160001.jpg image most likely belongs to i with a 99.99 percent confidence.
210705_161002.jpg image most likely belongs to i with a 100.00 percent confidence.

beranm14/parking