Vehware --- Ali Arda Eker
09/14/2016
- CPT form is prepared and submitted.
- Spoke with Mr. Bill (CEO of Vehware) about possible projects of his company which are about visual information processing.
- Spoke with Prof. Yin about his project with Vehware.
09/21/2016
- CPT form is accepted and initialized.
- Prof. Yin approved that I can work in his project with Vehware.
- Bill arranged a meeting for me to speak with a executive from Coca-Cola who is interested in image processing.
- Watched the demo made by Bill to the executive from Coca-Cola about visual information processing.
09/28/2016
- Talked to the Peng Liu who is the phd student of Prof. Yin. We will be working in the eye tracking project together.
- Installed related software and frame work to my pc. OpenCV and visual studio 2015 will be used.
- Sent an e mail to Peng. Waiting to hear from him or Prof. Yin about what is next.
10/05/2016
Goal of the project:
Creating a smart wheel chair that will capture the patient's eye motions to move wheel chair where he or she looks without the need of using hand gestures.
Requirements:
There will be cameras attached to the wheel chair to sense the patient's eye movements. Patient will not need to wear a helmet.
I am assigned to implement adaBoost algorithm for openCV. It is a supervised classification method that combines the performance of many week classifiers instead of a monolithic strong classifier such as SVM or Neural Network. Decision Trees are most popular week classifiers for Adaboost (Adaptive Boosting). Decision Tree means a binary tree (each non-leaf node has 2 child) and for classification, each leaf is marked with a class label. Multiple leaves may have the same label.
Adaboost Model:
Y = F(X)
X(i) ∈ R(k), Y(i) ∈ -1, +1
k = component vector. Each component encodes a vector.
Boosting Types:
1) Two-class Discrete Adaboost (what I will implement)
2) Real Adaboost
3) LogitBoost
4) Gentle Adaboost
10/12/2016
There are 2 main approaches for face detection.
- Image Based Methods: Use classifiers trained statically with a given example set. Then classifier is scanned through the whole image. Boosting technique will improve classifiers.
- Feature Based Merhods: Detecting particular face features as eyes, nose, etc.
fn = 1 - d
- Detection Rate (d): Percentage of faces in the image that have been correctly detected. We want to maximize this.
- False Negative (fn): Rate of faces that are forgetton.
- False Positive (fp): Rate of non faces windows that are classified as facesç We want to minimize this.
