Today, many autonomous vehicle systems use LIDAR or other technology to have some sense of depth. However, these technologies are often expensive, and a large bulk of the cost of an autonomous vehicle. We aimed to, using purely vision based techniques, avoid obstacles by object detection and also try to navigate on a road.
What it does
We used a jumping sumo on-the-ground drone that jumps over obstacles at the right time (not just when it sees them, but when it knows that it should jump over them -- which is a hard problem to program). We collected a lot of our own training data, and trained our own deep learning model, finetuning SSD, to create an object detection model, and we used object localization to determine when to jump over an object.
We also created a track for our robot and used more classical computer vision techniques to try and navigate this track.
How I built it
Challenges I ran into
Hardware is difficult to use! Lots of latency and networking related issues that made it hard to do in real time and just work with it.
Accomplishments that I'm proud of
What I learned
What's next for Autonomous Jumper