Research Paper link - https://drive.google.com/file/d/1eBG344O3m9ZIDXhHwdzXj6rjKtfqbKDS/view?usp=sharing
In my country, road-related fatalities are extremely widespread. They are often due to a lack of proper infrastructure and the poor quality of roads. In a bid to supplement those in need around me, and put my skills to the test, I undertook this research project. I learned everything during the course of the hackathon
What it does
We worked on a consumer-facing model ported on mobile devices to help reduce cycling accidents. Our iOS application is designed to detect potholes in real-time when mounted on a cycle with the camera facing the road.
We use SOTA algorithms for object detection and tracking to complete our software. The SSD MobileNet model is trained for over 30,000 images of 4 types of potholes. We integrated this SSD model into an app and then are in the process of using this app to register potholes for local municipal corporations. The model works with an mAP of 60% and is definitely going to benefit both the government with automatic road analysis and the general public with avoiding accidents. We used the TF converter to create an edge device compatible model and adapt real-time detections on a phone. The project can help governments significantly reduce accidents and better register pothole data.
How I built it
Aggregated data from the internet and around me. Wrote TF script and trained model. Learned iOS dev and ported it into IOS.
Challenges I ran into
With minimal knowledge of iOS, a significant challenge was a smooth port process from MobileNet into iOS. I usually work with YOLO so SSD was also a challenge.
Accomplishments that I'm proud of
Creating an all-encompassing system capable of helping municipal corporations, bikers, automated cars to add to their systems. Can be used on the road as well as for smooth collection of data. Also really proud of getting it published to app store. Note: It is published under my brother's dev account since we didn't have one.
What I learned
I learned a lot about Object Detection as well as iOS app development. It was an enriching experience
What's next for RoadSafety - Research
I would like to aggregate more data, improve my model and distribute it to people around me