TL;DR
We are now capable of reading in and detecting objects in both blind spots and the rear-view camera at 25 FPS @ 480p.
Big Software Steps
Cameras ✓
Improved FPS ✓
ML Detection ✓
Our software is just about done! Our blind spot camera feeds can swap to our rear-view feed smoothly. We're also hitting a respectable 25 FPS @ 720p on our dual blind spot camera feeds. This is with the ML model running so we can expect this to be our final FPS unless we decide to overclock or, if time permits, further optimize the model.
We've also reduced the model's responsibilities. Originally, it was detecting over 90 different objects - including chairs, bottles, and phones. We don't really need to detect these things while driving so we've cut the model down to only detect 5 key objects:
- Cars
- Trucks
- Pedestrians
- Bikes
- Busses
In the demo picture above, we show our dual cameras running simultaneously and feeding the feed into the model which detects our team member, Justin, as a pedestrian correctly.
What's Next?
We're looking to add a button to trigger the switch between rear-view and blind spot view as well as making some slight modifications to the 3D printed housings to better fit our parts.
Comments