alknemeyer / EEE4022S-Thesis-Project

Animal tracking using machine learning with limited computational power

EEE4022S Thesis Project: Autonomous Object Identification and Tracking using Neural Networks

Git repository to backup work for my fourth year thesis project at the University of Cape Town. The repo is mostly just to back up my work across multiple work stations, so a lot of the commit descriptions are horrible. Woops.

Abstract

Many systems require a camera to autonomously identify the objects in its field of view, and then rotate the camera to keep it aimed at a certain class of object.

This report describes the design and implementation of such a system on a low-powered device. The system uses a Convolutional Neural Network to identify the position of objects from photos, making use of a neural accelerator device to achieve near real-time inferences. These measurements are improved through the use of a Kalman Filter, which estimates the angular state of the tracked object. A multiprocess pipeline is used to manage the control on a non-real time OS. Finally, a gimbal is designed and built to rotate the unique payload.

The tracking system is shown to improve over a stationary camera setup.

Coding style:

I find development using a notebook to be quite a bit easier than developing using a regular python file. Unfortunately, you can't import a .ipynb as a module. So, here's the workflow:

1. Use each .ipynb file to understand the code and make changes.
2. When you want to commit a change, click Kernal > Restart and Clear Output to remove your outputs + make the file a bit smaller (shows up as fewer lines in the git commit).
3. Run the command jupyter nbconvert --to=python FILENAME.ipynb to generate a .py file which can be imported as a module. Just make sure that any debugging code doesn't run if this is imported as a module into another file!

Alexander Knemeyer