A kind of face tracking methods of multi-cam collaboration
Technical field
The present invention relates to technical field of face recognition, more specifically to a kind of face tracking methods.
Background technology
In existing face recognition technology, realized respectively and face detected, to the face that detects with
Track, face characteristic extraction is carried out to given image or video sequence and is compared with face database data so as to be carried out to user
Identification, but conventional face's identifying system is easily influenceed by the various external conditions such as illumination, beard, glasses, hair style, expression, makes knowledge
Rate does not reduce.Therefore, system availability is not strong.And the recognition of face Robust Algorithms of adaptive illumination variation can be in certain journey
Face identification rate is effectively improved on degree, but still still there is limitation, it is impossible to efficiently solves the influence of various scene changes.
In recognition of face, the discrimination of better quality facial image also can be higher, therefore, illumination estimation, Attitude estimation and fuzzy
The Quality estimation of the facial images such as estimation is the most important thing.
The content of the invention
The present invention is in order to overcome above mentioned problem, there is provided a kind of method of the face tracking of multi-cam collaboration, to improve
The efficiency and utilization rate of man face image acquiring, so as to improve the discrimination of face.
In order to realize above mentioned problem, present invention employs following technical proposal:
A kind of method of the face tracking of multi-cam collaboration, its specific steps include,
S1. area-of-interest delimited according to monitoring scene;
S2. multiple cameras are set in area-of-interest, distribution camera enables to cover area-of-interest;
S3. it is just right after any one camera detects a more complete face when personnel enter area-of-interest
This face is tracked using improved TLD track algorithms;
S4. after target face leaves from a certain camera picture, it is laid out according to camera in area-of-interest non-thread
Property graph structure, calculate camera field of view registration and camera in facial image similarity, determine graph structure search for
Weight;
S5. the camera according to side right weight from big to small, carries out Feature Points Matching to face and target face successively, until
Untill the match is successful, if one circulates, matching is all unsuccessful, then illustrates that this person has left area-of-interest;
S6. Quality estimation is carried out by picture quality optimal policy to all target facial images traced into, according to matter
Amount fraction selects out an optimal facial image of quality and is used for follow-up recognition of face.
Preferably, the improved TLD tracing algorithms in the step S3 detect a whole person in any camera
After face, this target face is constantly learnt, positive and negative samples storehouse is constantly updated, is detected for target location in subsequent frame.
Preferably, in the step S4 calculate camera field of view registration in camera facial image it is similar
Degree, is according to formula:Wi=α Ri+βSi, wherein 0≤α, β≤1, alpha+beta=1.
Preferably, the mass fraction in the step S6 is to image blur estimation, illumination estimation, Attitude estimation, glasses
Detection judges that five aspects give certain weight Wi respectively with form, is according to formula:scorei=w1fi+w2li+w3pi+w4gi+
w5mi。
Compared with prior art, the device have the advantages that:
Face tracking is carried out using more shootings, multiple different illumination can be obtained in the area-of-interest of monitoring scene,
The target person face image such as different postures, fuzziness difference.By setting certain weight, the facial image of multiple different qualities is entered
Row quality evaluation, select out an optimal facial image of quality.The effect of man face image acquiring is effectively improved by this scheme
Rate, the discrimination of face is substantially increased, there is very big practical value.
Brief description of the drawings
Fig. 1 is area-of-interest and camera position schematic diagram in embodiments of the invention;
Fig. 2 is TLD technical work principle figures;
Fig. 3 is Quality estimation flow chart in the embodiment of the present invention.
Embodiment
The present invention is further described with reference to the accompanying drawings and detailed description:
A kind of method of the face tracking of multi-cam collaboration, with reference to figure 2,1~9 is camera in figure, and A is region of interest
Domain, B are monitor area, then its specific steps includes,
S1. area-of-interest A delimited according to monitoring scene B;
S2. area-of-interest sets 5 cameras in the present embodiment, and distribution camera enables to cover region of interest
Domain;
S3. such as Fig. 2, when personnel enter area-of-interest, it is assumed that detect a more complete face by camera 1
Afterwards, just this face is tracked using improved TLD track algorithms;
S4. after target face leaves from a certain camera picture, it is laid out according to camera in area-of-interest non-thread
Property graph structure, calculate camera field of view registration and camera in facial image similarity, determine graph structure search for
Weight;
S5. the camera according to side right weight from big to small, carries out Feature Points Matching to face and target face successively, until
Untill the match is successful, if one circulates, matching is all unsuccessful, then illustrates that this person has left area-of-interest;
S6. Quality estimation is carried out by picture quality optimal policy to all target facial images traced into, according to matter
Amount fraction selects out an optimal facial image of quality and is used for follow-up recognition of face.
In the present embodiment, with reference to the fundamental diagram that figure 2 is TLD technologies, the improved TLD trackings in the step S3
After algorithm detects a complete face in any camera, this target face is constantly learnt, constantly updated positive and negative
Sample Storehouse, detected for target location in subsequent frame.TLD technologies include tracker, learner, detector, integrator and regarded
Frequency frame.The movable information that tracker is changed using frame to frame tracks target, and learner is assessed detector, and corrigendum avoids sending out
Raw same mistake.Frame of video is for gathering face characteristic, and integrator is that face is compared with target face characteristic.With
Track device mainly applies pyramid optical flow method, and learner is to complete online real-time learning by random forests algorithm.With reference to reality
Border scene, is improved and optimizes to Face tracking algorithm, and the robustness for improving Face tracking algorithm is very crucial, to follow-up
The Quality estimation of facial image has a major impact with recognition of face.
Preferably, in the step S4 calculate camera field of view registration in camera facial image it is similar
Degree, is according to formula:Wi=α Ri+βSi, wherein 0≤α, β≤1, alpha+beta=1, Wi is the weight of each edge in graph structure, RiTo take the photograph
As the registration of head field of view, SiFor the similarity of facial image in camera.It is laid out according to camera in area-of-interest
Non-linear graph structure, calculate camera field of view registration and camera in facial image similarity, it is determined that figure knot
The weight of structure search.According to the weight between each two camera being each edge, it is possible to preferably solve the association of multi-cam
The same sex.
Preferably, with reference to figure 3, the mass fraction in the step S6 is that image blur estimation, illumination estimation, posture are estimated
Meter, Glasses detection and form judge that five aspects give certain weight Wi respectively, are according to formula:
scorei=w1fi+w2li+w3pi+w4gi+w5mi
Wherein, scoreiFor mass fraction, fiFor image blur estimation, liFor illumination estimation, piFor Attitude estimation, giFor eye
Microscopy is surveyed, miJudge for form.
Embodiments of the invention are the foregoing is only, are not intended to limit the scope of the invention, it is every to utilize this hair
The equivalent structure or equivalent flow conversion that bright specification and accompanying drawing content are made, or directly or indirectly it is used in other related skills
Art field, is included within the scope of the present invention.