Gaze Detection Models, Features

Predicts when pedestrian is engaging with camera based on facial keypoints via binary classification.

Deploys browser site to livestream predictions via Flask (app.py for deployment from PC, ros_web_docker/pub_stream.py for deployment with ROS publishing).

Total Features List

• A. Nose-Eye Offset

  • Ratio of horizontal distance from eye center to nose and eye distance.

• B. Ear-Nose Offset

  • Ratio between perpendicular distance from earline to nose and earline length.
    • earline = [Left Ear -- Right Ear]

• C. Nose-Eyeline Ratio

  • Ratio between Pythagorean distances of [Left Eye -- Nose] and [Right Eye -- Nose].

• D. Number of Ears [0~2]

  • Number of ears confidently detected by YOLO.

BEST MODEL: Gradient Boosting 2param

• VA: 76%

• Features: A, B

  • Most consistent, does not flicker btwn APPROACH/DNI as much.

GB/RF NENl: Nose, Ear Offsets + Nose-Eyeline Ratio

• VA: 70%, 74% (GB, RF)

• Features: A, B, C

RF v77: Random Forest

• VA: 78%

• Features: A, B

  • Highest accuracy, but in practice worse; flickers often.