Road-Accountability

Crowdsourcing road maintenance accountability.

Comment

Inspiration

Potholes have over the last few years become more common on the roads that I use, and recently I’ve realised that the Government are more likely to fix these roads if somehow they could remotely detect them.

What it does

RoadAccountability is a simple web platform that would allow road users take picture of potholes they come across, and these potholes are checked by a deep learning model in the back to assert that they are indeed potholes, thereafter logging the pothole pictures and their locations to a map that can be seen by all.

How we built it

System Design:

  1. Trained a deep learning model from Road data on Azure Machine Learning.
  2. Deployed the model as a web service running in an Azure Container Instance.
  3. Built the web app locally with Django, from Django the model is called during image uploads.
  4. If an Image sent to the model is deemed to be a Pothole, the image's metadata is extracted, and the image is then stored with its geographical location and timestamp.
  5. Using folium python library, the potholes are populated as markers on the map.

Public Dataset : https://zindi.africa/competitions/miia-pothole-image-classification-challenge

Challenges we ran into

Web Development : Not having any previous web development experience, it was a challenge getting the platform up and running in the way envisioned.

Accomplishments that we're proud of

The system flow, the POC's working parts give a glimpse of how massively effective and scalable this could be, given its practicability.

What's next for Road-Accountability

  1. Tie the loose ends and have the platform working as intended, and deployed.
  2. Maintain and build together with the opensource community, the web development part of it is a whole lot and crucial to usability. Community only can deem this useful, effective and determine its success.
  3. Improve Model Accuracy, currently at 76%. However there is a concern with the distribution of images used to train the model vs images likely to be uploaded by users. The images used in training were from a dashcam, and users are more likely to take images from a closer point of view with cellphones.
Share this project:

Updates