Camera system for traffic enforcement

1. A traffic enforcement system, comprising:

a camera configured to capture an input image of one or more subjects in a motor vehicle;

a memory storing a deep learning model configured to perform multi-task learning for a pair of tasks including a liveness detection task and a face recognition task on one or more subjects in a motor vehicle depicted in the input image; and

a processor configured to apply the deep learning model to the input image to recognize an identity of the one or more subjects in the motor vehicle and a liveness of the one or more subjects, and wherein the liveness detection task is configured to evaluate a plurality of different distractor modalities corresponding to different physical spoofing materials to prevent face spoofing for the face recognition task.

2. The traffic enforcement system of claim 1 , wherein the processor is further configured to forward the input image of the one or more subjects in the motor vehicle to a corresponding traffic enforcement authority, responsive to the motor vehicle committing a traffic violation.

3. The traffic enforcement system of claim 1 , wherein the processor is further configured to forward the input image, that depicts a particular one of the one or more subjects that is located in an operating location in the motor vehicle, to a corresponding traffic enforcement authority, responsive to a liveness determination of another one of the one or more subjects in the motor vehicle being determined as not live and a location of the motor vehicle being in a high occupancy vehicle lane.

4. The traffic enforcement system of claim 1 , wherein the processor is further configured to forward the input image, that depicts a particular one of the one or more subjects that is located in an operating location in the motor vehicle and another one of the one or more subjects located in a non-operating position in the motor vehicle, to a corresponding traffic enforcement authority, responsive to a liveness determination of the another one of the one or more subjects being determined as not live and a location of the motor vehicle being in a high occupancy vehicle lane.

5. The traffic enforcement system of claim 1 , wherein the different physical spoofing materials comprise a spoofing materials selected from the group consisting of a 3D mask, a printed image, and an electronic display.

6. The traffic enforcement system of claim 1 , wherein the processor is further configured to perform data preprocessing on the input image selected from the group consisting of image contrast enhancements, data augmentation, and cropping.

7. The traffic enforcement system of claim 1 , wherein the processor is further configured to apply an image contrast enhancement to the input image by randomly adding one or more Gaussian distributed instances to an original RGB value of the of input image.

8. The traffic enforcement system of claim 1 , wherein the processor is further configured to apply an image contrast enhancement to the input image by adding one or more Gaussian distributed instances to a Hue, Saturation, and Value (HSV) domain and multiplying two or more coefficients in the HSV domain to further adjust the Saturation and the Value of the input image.

9. The traffic enforcement system of claim 1 , wherein the processor is further configured to perform random central cropping on the input image.

10. The traffic enforcement system of claim 1 , wherein the processor is further configured to perform a data augmentation process on the input image to augment data in the input image and increase a generality of the deep learning model to the different distractor modalities.

11. The traffic enforcement system of claim 1 , wherein the deep learning model is formed to include a plurality of layers, and the processor is further configured to train the deep learning model by freezing each of the plurality of layers except a last fully connected layer therein, and performing multi-task training by combining the liveness detection task and the face recognition task, wherein during the multi-task training, all weights of the plurality of layers are unfrozen.

12. The traffic enforcement system of claim 1 , wherein the deep learning model is formed using an inception structure configured to multiply an output of a convolution layer of the deep learning model with the input image to produce a new image having amplified extracted features relative to the input image.

13. The traffic enforcement system of claim 1 , wherein the deep learning model is formed using a loss structure that includes a face liveness penalty and a face recognition penalty.

14. A computer-implemented method for traffic enforcement, the method comprising:

capturing, by a camera, an input image of one or more subjects in a motor vehicle;

storing, in a memory, a deep learning model configured to perform multi-task learning for a pair of tasks including a liveness detection task and a face recognition task on one or more subjects in a motor vehicle depicted in the input image; and

applying, by a processor, the deep learning model to the input image to recognize an identity of the one or more subjects in the motor vehicle and a liveness of the one or more subjects, and wherein the liveness detection task is configured to evaluate a plurality of different distractor modalities corresponding to different physical spoofing materials to prevent face spoofing for the face recognition task.

15. The computer-implemented method of claim 14 , further comprising forwarding the input image of the one or more subjects in the motor vehicle to a corresponding traffic enforcement authority, responsive to the motor vehicle committing a traffic violation.

16. The computer-implemented method of claim 14 , further comprising forwarding the input image, that depicts a particular one of the one or more subjects that is located in an operating location in the motor vehicle, to a corresponding traffic enforcement authority, responsive to a liveness determination of another one of the one or more subjects in the motor vehicle being determined as not live and a location of the motor vehicle being in a high occupancy vehicle lane.

17. The computer-implemented method of claim 14 , further comprising forwarding the input image, that depicts a particular one of the one or more subjects that is located in an operating location in the motor vehicle and another one of the one or more subjects located in a non-operating position in the motor vehicle, to a corresponding traffic enforcement authority, responsive to a liveness deteimination of the another one of the one or more subjects being determined as not live and a location of the motor vehicle being in a high occupancy vehicle lane.

18. The computer-implemented method of claim 14 , further comprising performing data preprocessing on the input image selected from the group consisting of image contrast enhancements, data augmentation, and cropping.

19. The computer-implemented method of claim 14 , wherein the deep learning model is formed to include a plurality of layers, and the method further comprises training the deep learning model by:

freezing each of the plurality of layers except a last fully connected layer therein; and

performing multi-task training by combining the liveness detection task and the face recognition task, wherein during the multi-task training, all weights of the plurality of layers are unfrozen.

20. A computer program product for traffic enforcement, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:

capturing, by a camera, an input image of one or more subjects in a motor vehicle;

storing, in a memory, a deep learning model configured to perform multi-task learning for a pair of tasks including a liveness detection task and a face recognition task on one or more subjects in a motor vehicle depicted in the input image; and

applying, by a processor, the deep learning model to the input image to recognize an identity of the one or more subjects in the motor vehicle and a liveness of the one or more subjects, and wherein the liveness detection task is configured to evaluate a plurality of different distractor modalities corresponding to different physical spoofing materials to prevent face spoofing for the face recognition task.