At the Entrepreneurs First EBL Munich Hackathon, Janik, Alex, and I built a prototype for EduPin, a small device children can clip to the strap of their backpack that teaches them how to navigate traffic.
We finetuned a small YOLO to be able to detect red and green pedestrian traffic lights, crosswalks, and stop signs. Our training data was three public datasets we merged for this purpose. A pruned and quantized version of the model runs on a Raspberry Pi with AI camera and does live inference.
Once a dangerous or educative situation is detected (demo: presence of a red pedestrian traffic light) the device starts an educative dialogue to teach the child about the situation (demo: asking what to do when encountering a red traffic light and not stopping until hearing the right answer).
After some initial test runs, yoloing a “large” (for the ~30 hours we had) training run, and lots of trouble with the quantization tool, we were happy everything worked out at the end.
The jury seemed happy too, awarding us the hackathon win and EQT award. Thanks again to the EF team for organizing such a nice hackathon and to EQT and Cursor for the awesome prizes!
Our pitch again. (And my laptop being repurposed as a traffic light.)
