Adapt

What inspired your project?

According to the national highway traffic administration, car accidents happen nearly once a minute, amounting to over five million car accidents each year. With statistics like these, car accidents are extremely common and can happen to anyone. Just in the past year, multiple members of our group have been either victims or had family members fall victim to a car accident. Thankfully, no one was hurt. But, their families were left without access to a vehicle for over a month, and in a country so heavily reliant on the ownership of a motor vehicle, this was a massive hindrance to their daily activities.

What problem does your project solve?

It’s usually best to go through your insurance company for the repair of your motor vehicle. Insurance companies need a quote for the repair cost of a vehicle. However, this is an extremely inefficient process as providers send over agents to manually estimate the repair cost, a process that can take weeks or even months. The other alternative is to find a mechanic on your own and get their quote. This approach comes with two main issues. The first is that the consumer may not know whether or not a quote is fair, ie. a mechanic may say the repair cost is higher than it actually is in order to earn more money. The second is that without an evaluation from their insurance company, they won’t know how much of the repair cost their insurance will cover.

How does your project solve the problem?

Our project aims to solve this issue of getting an accurate quote for automotive repair cost with machine learning. After taking images of the damage to the car, the user can upload these images to our website. A machine learning algorithm then identifies the part of the car that is damaged and estimates the severity of the damage. These findings are then quantified into an accurate repair cost, providing a repair quote in a fraction of the time. This information can then be used in multiple ways. Insurance companies can use this tool to more rapidly tell the consumer how much insurance will cover, and consumers can use it to make better decisions about whether they should go through their insurance provider or seek out independent repair shops. In the case of major accidents, it may be better for the consumer to purchase a new car. For instance, if our model estimates that the cost of repair is over $10,000, the consumer may opt to buy a used car instead of trying to repair their current one. Overall, our project resolves a specific, automotive problem that persists in the space of personal finance: will my insurance cover my payment, and if not, should I bother getting my car fixed?

What technologies did your project use? Include programming languages, libraries, and any external tools.

Our project used the programming languages of HTML, JavaScript, and Python. Additionally, we utilized libraries such as TensorFlow, Keras, NumPy, and OpenCV. Lastly, we used external tools like Python OS, Google Colab, and GitHub. Specifically, HTML was used to create the website structure, JavaScript was used to provide functionality to the webpage, and Python was used for machine learning. With regards to libraries, TensorFlow and Keras were used for creation and training of the ML models, OpenCV was used for image processing, and NumPy helped with data processing. Finally, in terms of external tools, Python OS served as a medium to work with and open files, Colab was where we programmed the majority of our project, and GitHub was used to store repositories as well as for version control.

What challenges did you run into?

The very first struggle we had was finding a sufficient data set for the images of damaged cars, along with appropriate labeling. Due to time constraints, we didn’t want to manually label, so instead looked for a data set with predefined classifications and bounding boxes (around damaged areas). Fortunately, we were eventually able to find a data set that met our requirements. Secondly, we had issues with training the machine learning model due to multi output classification. Since we had five different areas or parts of the car that we focused on, the model would often act up in erratic ways that we didn’t understand. We solved this problem by creating a model for each of the five areas of the car instead. Thirdly, we ran into issues surrounding connectivity between the backend (ML model) and the frontend (a potential UI). We attempted to use Django to create an API that communicated information from the front end to the back end, but this ultimately proved to be extremely complicated.

Built With

• javascript
• keras
• numpy
• opencv
• python
• tensorflow