• 4x 60s 20hz video

• Annotated face tracking video
• Head pose feature vector

• numpy
• sklearn
• skimage
• cv2

1. Core

   • Frame preprocessing
   • Facial detection
   • Facial landmark identification
   • Geometric orientation
   • Rendering
   • Main

2. Support

   • Configuration
   • SVM preprocessing
   • SVM training
   • File utilities
   • Numpy utilities
   • Image utilities

3. Data

   • Trained SVM
   • Input
     • Video files
   • Intermediate
     • Preprocessing (optional)
     • imagesToFaces
     • videosToFrames
   • Output
     • Annotated video
     • Head pose estimation feature vectors
   • Haar cascade classifier

4. Spike

   • Random excursions

5. Tests

   • TBD

video -> video preprocess + dataset preprocess -> train svm -> face detection -> retrain svm -> face detection -> find landmarks -> calculate geometry -> render

read frame -> frame preprocess -> face detection -> find landmarks -> calculate geometry -> render

1. HEVC video dataset -> frames
2. Yale faces dataset -> cropped

1. Cropped yale faces -> positive samples
2. 256 object categories dataset -> negative samples
3. Annotate samples
4. Train linear SVM
5. Save SVM model
6. (After sliding window): Retain SVM with hard-negative mining
7. Save new SVM model

1. Sliding window over image pyramid
2. Non-maximum suppression

1. Facial landmark alignment
2. 2D-3D point mapping
3. Compute head orientation

1. Pupil detection
   • CDF
   • Feature Extraction and Normalization
2. Gaze Classification and Decision Pruning

Method:

Using comma ai dataset: Take in hevc video Extract frames from 60s of 20hz video (~1200)

Using yale faces dataset: Convert to jpg and grayscale Crop the images using builtin haar cascades uniform resize write to disk generate (~165) positive samples for SVM using skimage hog descriptor

Using 256_object_categories dataset: generate (~30600) negative samples for SVM using skimage hog in batches of 1000 saving to disk

arrange data correctly + add labels train svm with the positive and negative samples save trained svm

sliding window image pyramid non-maximum suppression hard negative mining retrain

find face find eyes geometric transformation for facial plane generate vector