ScratchMatch

• 

  original image

• 

  segmented car

• 

  segmented tires

• 

  roi rectangles

Inspiration

When you return the rental car to Sixth, there is always this fun moment that someone has to check if the car has been damaged during your drive. This means if you are the lucky person that has managed to come back in the working hours, you have to wait for a friendly employee to become available, drag him to the car, wait for him to check it and than you are likely to get a conformation that everything is fine. However, if you are back from the late drive, as we normally are, you might just end up dropping the car off without any idea when you will get the information that everything was fine. Hoping that nothing will happen until the morning...

What it does

We want to address this problem by building automated detector for scretches on the rental car. With the static cameras installed on one parking lot slot, the customer could drive in certain position and get his/hers car scanned for demages and receive conformation on the spot if there are any new one on this vehicle. For the prof of consept we have decided to play with the small Sixth cars, which were drawing our attention sitting so lonely in MI Magistrale. We have a static laptop camera which scans the Sixth toy car while pooling into the parking spot from behind. It will detect relevant area representing the vehicle surface, compare it to previous state which has been stored in the memory and display if there are any new demages to the car. Once these damages have been found, you can simply confirm them with the button press after which point the model will get updated and these damages will be displayed as already familiar, old one on the car. The algorithm will always look for new damages.

How we built it

We have build full project with just the use of Python and openCV library for image recognition. We are taking live video feed from computer camera and first extracting roi (region of interest) which will in this case be the surface of the car visible from this camera angle. After this we analyse the car surface with canny edge detector in order to find lines and egdes in the image. They are saved to the "heat map" which will detect the areas of old and new damages.

Challenges we ran into

Computer Vision algorithms as fun as they might be are not so grateful for fast and rubust implementation. The biggest problem for our challange was to make this proof of consept run stable.

Accomplishments that we're proud of

Rear moments of creativity:

No more stres
with rental car mess
just drive to your new camera spot
that Sixth has on each parking lot
and let our automated scratch detection
tell if car is in the same state of perfection!

What we learned

Even the small toy Sixth cars are not always easy to drive. ;)

What's next for ScratchMatch

Transfering the idea to the real-world use case would enable the use of better quality cameras for more precision. The more general car damage detection should be done with training the neural network model to recognise these kind of damages on the car as well as car regions more stable.

Built With

• opencv
• python