A project within the class "ECE 740 - Computer and Robot Vision" at University of Alberta.
Shane Allan, Sophia Wagner, Almas Sahar
In this work, an enhanced depth estimation methodology is proposed using a low-cost stereo-vision camera that can deliver quality depth images in low-light environments suitable for real-time implementation. To address this problem, a convolutional neural network (CNN) is recommended to enhance the raw stereo red, green, and blue (RGB) and depth images for use on a prototype unmanned aerial vehicle (UAV). Our contributions are the following:
- Development of different CNN models in order to enhance depth from low-light depth and stereo images.
- Generation of a dataset with varying quality and lighting conditions consisting of depth and stereo images.
- Running the CNN models on the onboard computing module of the UAV.
The generated dataset can be found here
Software:
- PyTorch 1.10.0
- Python 2.7, 3.6
- C++
- ROS Melodic
- Ubuntu 18.04
Hardware:
- DJI F450 Flamewheel Frame
- Garmin ToF Laser Rangefinder
- Pixhawk 5x Flight Controller
- Sereolabs Zed 2 Camera
- Jetson Xavier NX
| Architecture | Results |
|---|---|
 |
 |
| Architecture | Results |
|---|---|
 |
 |