Use sensors on road intersections to optimize traffic lights, using deep learning

Project by Bruno

Abstract

Notraffic uses sensors on road intersections to detect road users (6 classes in total) to control the traffic lights in an optimized way. So in their use case is more important not to miss a road user than getting the class right. Therefore, the goal of the project was to make the model consider that use case and perform better in that sense.

Challenges

The detector is implemented by using TensorFlow Framework, and the first task was to dig deep into the code and configuration and understand the architecture and training process of the model.

After understanding the code, I found out what parameter I would have to change, but in order to tweak it right, I needed to understand the values it can assume. So to do that, I’ve used a tool to generate bounding box and developed a code to report the values of the example I build in the box generator tool. 

Achievements (according to KPIs)

  • 2 new customized loss functions
  • The capability of evaluating an experiment as class agnostic at the same time as evaluating normally
  • Developed a research tool to understand loss values and be able to tweak it 

Further development 

Use the developed custom loss functions and play with the numbers

Supervisor Feedback

Bruno was a deep learning intern under my guidance for the past 6 weeks. He designed and implemented several custom loss functions that help take a standard object detector and better fit it to the company’s use case.

Bruno was able to research and understand the state of the art computer vision models down to the last detail. In a very short period of time, he understood the major gaps between research papers and their open-source implementation and was able to bridge that gap with great coding abilities. He wasn’t afraid to dig deep into Tensoflow code and into the unstructured and undocumented world of open source.

Bruno went above and beyond and added tests to every element he added to our repository. Bruno was very keen on learning while challenging me with great questions that show a thorough understanding.

With a very strong background in software development combined with the ability to research and apply complex deep learning concepts, Bruno’s achievements were far above my expectations. I would have like to have him join our team immediately and he would be the first candidate I would reach out to once our startup grows.

 

 – Yoav Valinsky, Computer Vision researcher at NoTraffic

Share this post

Share on facebook
Share on twitter
Share on linkedin
Share on email