The objective of this project is to develop a UAV Vision system that can detect and avoid objects.
Sponsored by: Boeing
Team Members
Noah Dever Petr Esakov Trey Fishburn Shardul Jagtap Jenny Lynn Kelly Evan Soisson
Instructor: David Cubanski
Project Poster
Click on any image to enlarge.
Project Video
Project Summary
Overview
Boeing is at the forefront of innovation in the aerospace industry, exploring new ways to improve safety and efficiency. One area of focus is autonomous technology, which can enhance the predictability and reliability of products. In this project, we developed technology for a UAV that uses a machine-learning, camera-based, object detection system to detect and avoid obstacles. This technology has potential applications for unmanned package delivery and electric Vertical Takeoff and Landing air taxis, which are fast, efficient, and environmentally friendly transportation solutions.
Objectives
– Integration of machine learning-trained camera system onto the drone.
– Provide the pilot with real-time advice on object avoidance.
– Detect failures within the system and notify the pilot.
Approach
– Define sensor space and type of sensors to be used for machine learning-trained camera system integration onto drone
– Train the algorithm using an Off-the-Shelf imagery database of example objects and prioritize object detection over classification
– Define the required level of certainty for object detection and classification, as well as the number of classes
– Define the test environment, which includes trees, cars, people, and other obstacles
– Estimate the object state (size, position, velocity vector) and provide a confidence metric of object detection and classification
– Evaluate the completeness of the training and explore if there is a need to react differently to objects depending on their classification
– Demonstrate detection/annunciation of object in flight and develop a proposed control loop for reacting to the detected obstacle
– Develop a safety monitor for detecting failures within the system, with emphasis on vision system
– Define levels of autonomy, provide cueing/advice on obstacle detection, and develop human oversight of system actions with limited authority
– Conduct flight demonstration to test the effectiveness of the system
Outcomes
– Object detection has an accuracy of 92%
– Created a detailed instruction manual for future teams to use.