Yolo Object tracking camera system

Deep learning, Yolo, Algorithm, Arduino, Programming, darkent, Image resignation

Comment 2

Inspiration

The idea came from the sentry robot from portal game. I always want's a working sentry robot.

What it does

It ran a "cutting-edge" image regeneration algorithm called Yolo( You only look once) through the neural network ——Darknet. By using trained weights, It can achieve real-time object identification even on a laptop. By using such tech, we send the information to an Arduino that connected to a servo station to tracking the needed object.

How we built it

Darknet and yolo is open source code on GitHub, we started by building up the environment for the yolov2 algorithm then modify the code file to get the information for image tracking.

Challenges we ran into

There were lots of challenges 1.Build the environment for the neural network was not easy, I have to use very careful package and version control to make sure this algorithm running smoothly. It took lots of try and error and start over. 2.The lack of proper devices is also a challenge because we don't have a real webcam, I have to work with PS eye, which will not work well with Opencv. So we have to work around it.

  1. The lack of information online was also challenging. Because this is a really new algorithm (2016), not many people using it for a project, So many bugs and wire things don't have answer online.
  2. The lack of computing power is also very hard to work with. This algorithm usually run a research station that has a least a GTX titan X for computing. So when runing on laptop, it only can run around 6 fps max.
  3. Because we have to work with python, we have to use pyserial to transfer the data to Arduino and that was really had to work with because every bit needs are encoded with ASCII. That greatly limits our accuracy of the active tracking. 6.....

Accomplishments that we're proud of

We modify that code to achieve a really difficult function: output the location of the tracked object at realtime We all did something we never learned before.

What we learned

Never give up.

What's next for Person&Objects tacking with deep learning algorithms

Find a way to send the float point data without using the pyserial. That will make the tracking become super accurate. Its project will be amazing at security. As it can tell what passed through and can follow any suspicious object automatically.

Built With

Share this project:

Updates