This is a project that implements real-time object detection with custom and collected data on a Tello drone.
I used the official YOLOv5 model by Ultralytics and trained it on two custom datasets. I used a custom personal dataset with ~70 images annotated and curated, and one dataset Trash Detection Image Dataset was used to train a more accurate model with ~1800 images.
The personal model was labelled and processed using Roboflow.
My custom model was trained on approximately 80 images over 100 epochs with a precision of 64%, using YOLOv5s. Here’s a glimpse of its real-time detection performance::
The Roboflow dataset contains approximately 1,800 images and was trained for 50 epochs with a batch size of 244 using the YOLOv5s model with a precision of 90%! Here are the results produced by the model:
Here is a comparison between my personal dataset and the online roboflow:
This project uses real-time detection and logging for each run that you do. Each log is documented in the '/Logs' directory with each run being labeled '/runX' with X being the current run iteration.
In every '/runX' folder there is a corresponding .csv file containing logs of detections with above 90% confidence. If it detects an object with >90%, the confidence and time will be logged in the corresponding runX.csv file. Along with the logging in the file there will also be a corresponding screen capture logged in the '/runX/images' folder.
This is a python based project. You need python installed on your machine to run this program.
Must connect to a Tello drone and run python main.py in the shell. A black box will appear which is used to capture user input for controls. The drone can be controlled via the keyboard with the following commands:
- 'e': takeoff
- 'q': land
- 'w': forward
- 'a': left
- 's': backwards
- 'd': right
- 'K_UP': up
- 'K_DOWN': down
- 'K_LEFT': rotate ccw
- 'K_RIGHT': rotate cw