2. What is our project?
- Develop a drone that is capable of flying predefined routines for different
parking lots, in order to check vehicle parking permits.
4. Haar Feature-based Cascade
Classifier
”Rapid Object Detection using a Boosted Cascade of Simple Features” by Paul Viola and michael jones
-Train need positive/negative image
Negative Images Positive Images
-Use about 6000 negative images from imageNet(Stanford)
-Create about 3000 positive images using openCV from 6 original images
8. SVM (Support Vector Machine)
● Built SVM model(Using linear kernel) in python
● Train with ”MNIST” data (1 - 3 only)
○ 28 x 28 size handwritten digit image
○ Total data size : 21000
○ Training(15000)/Validation(3000)/Testing(3000)
○ Model testing using validation set
● About 90% accuracy on testing data set
MNIST Data
9. Digit Recognition
1. Crop
2. Pre-Processing the image
3. Rotate the image
4. resize
(28x28 MNIST Train data)
5. Read the digit using SVM
model
17. Flight part demo
Routine:
-Gets the list of coordinates.
-Goes to coordinates and takes photos
of the cars.
-Goes back to first point and lands.
//Video can go here
19. New Character Recognition Approach
-New method for extracting ‘Digit’ from detected sticker for better performance
-Add License detecting feature on application
-Fully understand visual pre-processing in License detecting part
20. Decoding Drone Camera Frames into
Bitmap
Drone Camera Frame
H_264 Encoded
Decode using
FFMPEG to
Bitmap
Input Bitmap image
Process Bitmap
image frame,
find object.
Processed Frame
Display on Android
View
Query whether
object was
detected or not.
21. Thank you!
For more of our project you can go to our blog where we have more videos of
flight testing, machine learning code and detailed explanation of what we do.
http://mightbesky.net/
Questions? Comments?