SeeingMachines FaceAPI -- real-Time 人臉追蹤工具

(資料來源reference)官方網址: http://www.seeingmachines.com/index.htm

採用即時的方式追蹤人臉

以上是官方的demo影片

  

Specifications

HeadTrackerV1	HeadTrackerV2	Notes
Tracking State			Tracking a face at framerate
Min Face Size (pixels)	40
Max Face Occlusion (%)	50
Head Rotation X (deg)	-20<45	-30<60	Horizontal axis (ear to ear)
Head Rotation Y (deg)	-30<30	-90<90	Vertical axis (up through head)
Head Rotation Z (deg)	-90<90		Camera axis (nose)
Positional Error (cm)	<1cm		Even illumination, no occlusion
Rotational Error (deg)	<3>		Even illumination, no occlusion
CPU Load (30hz USB webcam) (Total% / Process%)	12%/5%	30%/15%	Intel Core-2 Duo, 2.4GHz, 4MB Cache, Logitech Quickcam Deluxe for Notebooks
CPU Load (60hz firewire)  (Total% / Process%)	6%/1%	8%/1%	Intel Core-2 Duo, 2.4GHz, 4MB Cache, PointGrey Flea
Initializing State			Finding a new face
Min Face Size (pixels)	40.00		Distance between outer eye corners
Head Rotation X (deg)	<15		Horizontal axis (ear to ear)
Head Rotation Y (deg)	<15		Vertical axis (up through head)
Head Rotation Z (deg)	<30		Camera axis (nose)
Time to Acquire (typical) (secs)	0.3 - 3.0		Assuming head is within in geometric constraints
CPU Load while acquiring face	50%		Intel Core-2 Duo, 2.4GHz, 4MB Cache
Searching State			Quick recovery from tracking failure
Recovery Conditions	Face front, no occlusion	Any pose, no occlusion
Min Recovery Time	1 frame
CPU Load (%)	20%		Intel Core-2 Duo, 2.4GHz, 4MB Cache

官方的產品特色

Features

• Highly robust, real-time, 6 degree-of-freedom (3D) monocular face tracking

• Able to track up to +/- 90 degrees of head rotation and fast head movements

• Optional real-time tracking of lips and eyebrows

• Face-texture ”mugshot” delivery upon tracking commencement

• Two types of head-trackers, trading CPU load for tracking capability

• Requires only 8-bit greyscale video input. Algorithms do not rely on color and are therefore able to track faces in the

dark using infra-red illumination

• Automatic tracking startup and immediate reacquisition when face is hidden then shown

• Facial feature detection and tracking – locates facial feature “landmarks” from a single image or real-time for video

sequences

• Designed for all human faces (it doesn’t matter what you look like). Robust to occlusion, fast movement, large head

rotations, lighting changes, dimly lit rooms, facial deformation, skin color, beards, glasses etc

• Can track with as few as 40 pixels across the face (typically 2m from a VGA camera)

• Works with any webcam or video file

• Full control over all tracking parameters for purposes of software integration

• Low-level image input interface for custom camera integration

• “Offline” tracking algorithm for high-quality tracking of faces in movie files where real-time performance is not required

• Separate low-level “face-search” algorithm for finding multiple faces in images and short image sequences. This algorithm

has three “depth” levels, with level 0 providing fast results to level 2 which accurately locates facial features and estimates

head-pose in 3D