Yedidim

Feature Detection of Wheels

Data Science for Social Good

Margo Levin | Eli Bogdanov | Yaakov Haiby | Esther Berestetsky�

The Problem

When there is a tire with a puncture, a volunteer arrives and replaces it if necessary. ��

​

The Problem

​

When there is a tire with a puncture, a volunteer arrives and replaces it if necessary. ��

​

CALL

YEDIDIM

The Problem

​

When there is a tire with a puncture, a volunteer arrives and replaces it if necessary. ��

​

CALL

YEDIDIM

The Problem

​

When there is a tire with a puncture, a volunteer arrives and replaces it if necessary. ��

​

CALL

YEDIDIM

How to choose the right wheel?

The Problem

​

1. Junt diameter �

​

The Problem

​

1. Junt diameter �2. Center bore diameter�

​

The Problem

1. Junt diameter �2. Center bore diameter�3. PCD holes number �

​

The Problem

1. Junt diameter �2. Center bore diameter�3. PCD holes number �4. PCD diameter

�

​

The Problem

These features need to be identical in the replaced tire.

In the warehouse, there are many potential spare tires for replacement, but there is no tracking of these 4 features for each one.

​

The goal

Build an application that receives an image and identifies these 4 features

​

1. Junt diameter: 17 (inch)�2. Center bore diameter: 67.1 (mm)�3. PCD holes number: 5�4. PCD diameter: 114.3 (mm)

Strategy

1. Find junt, core bore, pcd, �bolt holes

​

Strategy

1. Find junt, core bore, pcd, �bolt holes

? HOW TO TRANSLATE �PIXELS TO INCH

Strategy

1. Find junt, core bore, pcd, �bolt holes
2. Find a reference object �which is of known size �(credit card, QR code)

Strategy

1. Find junt, core bore, pcd, �bolt holes
2. Find a reference object �which is of known size �(credit card, QR code)
3. Pixels-> inch/mm

1. Number of holes: 4

2. Junt diameter: 14.34 in

3. PCD diameter: 96 mm

4. CB diameter: 51.54 mm

Strategy

1. Find junt, core bore, pcd, �bolt holes
2. Find a reference object �which is of known size �(credit card, QR code)
3. Pixels-> inch/mm
4. Post processing

1. Number of holes: 4

2. Junt diameter: 14 in

3. PCD diameter: 100 mm

4. CB diameter: 54 mm

Strategy

Overall process:

1. Someone has a puncture
2. A volunteer comes
3. He sends the 4 features of the punctured wheel to a Telegram bot
4. The bot return True/False for having a valid replacing tire in the warehouse, and a picture of it

​

Main challenges

​

We faced many challenges throughout the process, including:

​

Data

- Given the image of the wheel, how to get it's true parameters from the internet?

​

Feature Detection

- Robustness of Segment Anything�- Images in bad perspective �- Bad quality images, images with�shadows

​

Card Detection

- Surfaces with rectangular patterns�- Little error = Huge error

​

Main challenges

​

And more…

​

• How to identify the reference object?
• How to convert from pixels to inches/mm?

• How to identify circles in the image?
• How to choose which circles represent the junt, the pcd, the core bore?
• How to deal with shadows?

Challenge Example: Pixel to Inch

We apply cv2.minAreaRect() to find the smallest fitting rectangle that contains the contour

Problematic Wheels

Add wheels here

​

Results

​

We build a Telegram Bot!

​

Credit Card

​

QR

Evaluation

• Percentage of success in holes number prediction: 94.12%
• We evaluated the predicted relations between features with ground truth on images of good perspective. The results are shown below:

Example of outliers

​

What’s next?

​

1. Theoretically, the punctured wheel’s features can be identified as well using a photograph, no need to manually type the features.
2. Also photos taken from a bad perspective can be processed
3. The Telegram bot can replaced with a dedicated application
4. Usage of QR Codes

Questions?

Thank you!