Vision based target tracking that distinguishes facial feature targets

1. A facial recognition system, comprising:

a face pair module connected to a neural network and configured to generate face pairs including positive and negative face pairs;

a multiple face tracking module configured to receive the face pairs from the face pair module and construct a trajectory model for identified human face; and

a fine tuning module connected between the neural network and the multiple face tracking module and configured to adaptively extract discriminative face features of the identified human face.

2. The facial recognition system of claim 1 , further comprising a pre- trained convolutional neural network; and

wherein the face pair module being connected to a neural network comprises the face pair module being connected to the pre-trained convolutional neural network; and

wherein the fine tuning module being connected between the neural network and the multiple face tracking module comprises the fine tuning module being connected between the pre-trained convolutional neural network and the multiple face tracking module.

3. The facial recognition system of claim 1 , wherein the multiple face tracking module being configured to construct a trajectory model for an identified human face comprises the multiple face tracking module being configured to link tracklets associated the identified human face.

4. The facial recognition system of claim 1 , wherein the multiple face tracking module being configured to construct a trajectory model for an identified human face comprises the multiple face tracking module being configured to construct the trajectory model using a pairwise Markov Random Field (MRF) model.

5. The facial recognition system of claim 1 , wherein the multiple face tracking module is further configured to:

derive a trajectory for the identified human face using the trajectory model; and

provide a person identity associated with the derived trajectory.

6. The facial recognition system of claim 1 , further comprising a face tracklet module configured to:

generate constraints indicative of: faces in a single tracklet being the same person and faces in different positions in a same frame being different persons; and

provide the constraints to the face pair module.

7. A facial recognition system, comprising:

a face tracklet module configured to:

form a face tracklet from a video frame;

generate spatio-temporal constraints indicative of: faces in the face tracklet being the same person and faces in different positions in the frame being different persons; and

provide the spatio-temporal constraints to a face pair module;

the face pair module connected to a neural network and configured to generate face pairs from the spatio-temporal constraints including positive and negative face pairs; and

a multiple face tracking module configured to receive face pairs from the face pair module and construct a trajectory model for an identified human face.

8. The facial recognition system of claim 7 , further comprising a pre-trained convolutional neural network; and

wherein the face pair module being connected to a neural network comprises the face pair module being connected to the pre-trained convolutional neural network.

9. The facial recognition system of claim 7 , further comprising a fine tuning module connected between the neural network and the multiple face tracking module and configured to adaptively extract discriminative face features of the identified human face.

10. The facial recognition system of claim 7 , wherein the multiple face tracking module being configured to construct a trajectory model for an identified human face comprises the multiple face tracking module being configured to link tracklets associated the identified human face.

11. The facial recognition system of claim 7 , wherein the multiple face tracking module being configured to construct a trajectory model for an identified human face comprises the multiple face tracking module being configured to construct the trajectory model using a pairwise Markov Random Field (MRF) model.

12. The facial recognition system of claim 7 , wherein the multiple face tracking module is further configured to

derive a trajectory for the identified human face using the trajectory model; and

provide a person identity associated with the derived trajectory.

13. A method comprising:

accessing a video frame;

forming a face tracklet from the video frame;

generating spatio-temporal constraints indicative of: faces in the face tracklet being the same person and faces in different positions in the frame being different persons;

deriving face pairs from the spatio-temporal constraints including positive and negative face pairs; and

constructing a trajectory model for an identified human face.

14. The method of claim 13 , further comprising adaptively extracting discriminative face features of the identified human face.

15. The method of claim 13 , wherein constructing a trajectory model comprises constructing a trajectory model using a pairwise Markov Random Field (MRF) model.

16. The method of claim 15 , wherein using the Markov Random Field (MRF) model comprises solving a joint object function and maximal condition probability of Markov Random Field (MRF) model.

17. The method of claim 13 , further comprsing:

deriving a trajectory for the identified human face using the trajectory model; and

providing a person identity associated with the derived trajectory.

18. The method of claim 17 , providing a person identity comprises providing the person identity using a Loopy Belief Propagation (LBP) algorithm.

19. The method of claim 17 , wherein deriving a trajectory for the identified human face comprises:

forming another face tracklet from another video frame; and

linking the face tracklet and the other face tracklet into the trajectory in accordance with the trajectory model.

20. The method of claim 19 , wherein accessing a video frame comprises access a video frame from a video sequence; and

further comprising accessing the other video frame from the video sequence.