Autonomous Vehicle Processor - Dec1710
The purpose of this project is to utilize an onboard system in realtime to provide information to flight systems regarding location of objects in a drone camera’s view. To accomplish this we use a pre-trained neural network to identify complicated objects and then a series of other, simpler methods to extract additional information such as distance away from the drone.
This software will utilize a system’s GPU to be able to more accurately and efficiently apply deep learning concepts to embedded real time system processing. It will be able to detect and analyze objects and provide relevant information about it.
The use of a neural network operation with real time data on onboard hardware has a wide range of applications beyond this project. While many military applications come to mind, there are plenty of civil uses as well to better the world. A good example of this would something like Amazon’s drone delivery service. If they could utilize this system, it could save time and money for the company by avoiding objects and ensure more happy customers by putting packages in more convenient and secure locations.
When finished, we want our system to be able to be installed onto any autonomous system and be able to provide accurate and useful information about the surroundings. This is intended for use on remote controlled drones, but this could also be applied to autonomous vehicles to detect objects near the vehicle or surveillance systems to detect intruders. The ability to detect a generic object has a multitude of uses already, but our system will be able to be trained for specific purposes. Additionally, the network to easily be trained to detect novel objects.
Lastly, here is a list of tentative features that may or may not be implemented:
- Distance and object orientation calculations
- Object trajectory predictions
- Simulation data collection
- Path correction
- Groundlink data feed
- Geographical position detection based on landmarks
- Multi Camera and alternative camera support
- Wingmate and other aerial object tracking