Devon Super

Computer vision is composed of the use of machine learning to interpret images and video. Computer vision algorithms are frequently developed to perform well on modern powerful computers, but these techniques are seldom designed to be deployed directly onto drones or other robots. Drones use computer vision machine learning techniques to navigate their environments. The goal of this project was to develop a pipeline that deploys deep neural networks onto drones that use a prominent AI execution module. The primary tool that makes up the pipeline is a program to execute deep neural networks on an industry-standard robotics framework. The pipeline also included a set of tools used to convert prototype depth estimation networks into efficient drone formats. The accuracy of the pipeline was tested using a state-of-the-art depth estimation network, and the node pipeline was able to run the network with the same accuracy achieved on desktop computers. GPU usage was not successfully enabled, so performance is not yet high enough to be used in real life applications. When GPU usage is present, based on previous performance tests, the expectation is that the pipeline will run the deep neural network at a speed that supports real-time usage. This pipeline bridges the gap between modern computer science research and real-life deployment by migrating networks from powerful computers, to the computers used by real-life drone applications.

Terra NYC STEM Fair Finalist, 2023

My Research Paper