Profile Identification
1. Objectives
- Automatic recognition for access control
- Remote control (no access card)
- Reduction of database search
2. Hypotheses
- Good record conditions: light, background, distance
- Cooperation: perpendicular to camera, proper distance
- Human supervisor for doubtful cases
3. Method
A) Contour extraction
Easy thanks to appropriate background
- Lowpass to reduce noise
- Hipass to enhance edges
- Region growing from background
- Border following of background region
B) Obvious reference points
- Top and Throat
- Chin and Nose
- Unfortunately, nothing reliable at the back
4. Feature extraction
Global
- Average width
- Back hair curvature
Chin-Nose
- 4 curvature maxima
- 3 distances and their sum
- Area
Chin-Throat
Nose-Forehead
- Not yet done (!spectacles)
5. Classification
Look for most similar feature vectors
- Nearest Neighbour (weighted L1-norm)
- Class score ordering
- Decision conditions (threshold, immunity with 2nd match)
Limitations
- Relative feature importance
- Independency
- Unrealistic values
6. Results
20 persons, 81 images (256x256)
immunity 1.0 1.1 1.2
recogn 88% 78% 68%
20 persons, 66 images (512x512)
immunity 1.0 1.1 1.2
recogn 98% 96% 83%
7. Further Topics
Front View / Profile combination
Model Matching
Obvious features: color, height, beard
3-D acquisition