Deep Learning: A Dive Into Autonomous Driving
School Name
Governor's School for Science & Mathematics
Grade Level
12th Grade
Presentation Topic
Engineering
Presentation Type
Mentored
Written Paper Award
2nd Place
Abstract
The Deep Orange project is working towards providing autonomous driving capabilities for vehicles, which has become a crucial function in the automotive industry. An autonomous steering function was created and demonstrated on a RC Car. An end-to-end solution utilizes a convolutional neural network (CNN) to analyse photos taken from a webcam mounted on the front and produces an output that directly correlates to a steering command. The end-to-end process proved very successful in driving the RC Car with zero user input. Training was performed using images taken on a scaled test track created specifically for this project. The individual images were correlated with a simple steering command (either left, straight, or right). The network was never explicitly told anything except for what direction the car was travelling, but was still able to successfully differentiate various conditions and elements that signified a turn or a straightaway. Instead of separating the process into different modules, such as lane detection and situational awareness, the end-to-end process provides an easy to implement system for automated driving. This solution provides a more affordable and less time-consuming process for implementing autonomous functions, due to its simplicity and absence of interfacing. In the future the system can be expanded to enable autonomous capabilities on vehicle prototypes and consumer vehicles.
Recommended Citation
Ho, Isaiah, "Deep Learning: A Dive Into Autonomous Driving" (2017). South Carolina Junior Academy of Science. 95.
https://scholarexchange.furman.edu/scjas/2017/all/95
Start Date
3-25-2017 11:59 PM
Presentation Format
Written Only
Group Project
No
Deep Learning: A Dive Into Autonomous Driving
The Deep Orange project is working towards providing autonomous driving capabilities for vehicles, which has become a crucial function in the automotive industry. An autonomous steering function was created and demonstrated on a RC Car. An end-to-end solution utilizes a convolutional neural network (CNN) to analyse photos taken from a webcam mounted on the front and produces an output that directly correlates to a steering command. The end-to-end process proved very successful in driving the RC Car with zero user input. Training was performed using images taken on a scaled test track created specifically for this project. The individual images were correlated with a simple steering command (either left, straight, or right). The network was never explicitly told anything except for what direction the car was travelling, but was still able to successfully differentiate various conditions and elements that signified a turn or a straightaway. Instead of separating the process into different modules, such as lane detection and situational awareness, the end-to-end process provides an easy to implement system for automated driving. This solution provides a more affordable and less time-consuming process for implementing autonomous functions, due to its simplicity and absence of interfacing. In the future the system can be expanded to enable autonomous capabilities on vehicle prototypes and consumer vehicles.
Mentor
Mentor: Andrej Ivanco, Clemson University International Center for Automotive Research