Real Time Face Tracking with pose estimation on tv clips - OPENCV

http://sites.google.com/site/ferrariimageprocessing/ : This is my face tracker with pose estimation tested on tv clips. It's developed with OpenCV using CONDENSATION algorithm (Condensation - conditional density propagation for visual tracking M. Isard and A. Blake) and the Haar - Adaboost cascade face detector (Robust Real-time Object Detection P.Viola and M. Jones). It's able to run until 60-70 fps with 320 x 240 images, on a commercial pc (2.2 Ghz, dual core).

Category:

Science & Technology

Tags:

• OPENCV
• FACE
• TRACKING
• POSE
• ESTIMATION
• DETECTOR
• VIOLA
• JONES
• CONDENSATION
• KALMAN
• FILTER
• REAL
• TIME

[0penCV]

With the last release it runs until 60 fps with a single core centrino 1.6Ghz and the false positives are in reality lacking. Enabling the pose estimation the performance decrease to 25 fps.

[0penCV]

the number of false positives is significant but this is due to the characteristic of the detectors (i did not specifically re-trained the classifier). Viola and Jones classifiers are known to trade-off efficiency for accuracy.

I've changed a lot of original source code so i have no longer cvHaarDetectObjects() method like OpenCV. Just the starting detections are with OpenCV cvHaarDetectObjects method.

[piechulla1966]

You get much too much false positives (the knees!) from the Viola/Jones-Detector. How does your argument list for cvHaarDetectObjects() look like?