System and method for training object classifier by machine learning
A system and method for training a computer-implemented object classifier includes detecting a foreground visual object within a sub-region of a scene, determining a background model of the sub-region of the scene, the background model representing the sub-region when any foreground visual object is absent from that sub-region, and training the object classifier by computer-implemented machine learning using the background model of the sub-region as a negative training example.
1. A method for training a computer-implemented object classifier, the method comprising:
detecting a foreground visual object within a sub-region of a scene within a field of view of a video capture device;
determining a background model of the sub-region of the scene, the background model representing the sub-region when any foreground visual object is absent therefrom; and
training the object classifier by computer-implemented machine learning using the background model of the sub-region as a first negative training example,
wherein the object classifier is trained specifically for a current scene, and wherein upon the current scene being changed to a new scene:
reverting to the object classifier without the training specific to the current scene; and
training the object classifier by machine learning using background models from the new scene.
2. The method of claim 1 , further comprising training the object classifier by machine learning using the detected foreground visual object as a positive training example.
3. The method of claim 1 , wherein determining the background model of the sub-region of the scene comprises:
selecting a historical image frame captured when any foreground object is absent from a sub-region of the historical image frame corresponding to the sub-region of the scene; and
cropping from the historical image frame the sub-region corresponding to the sub-region of the scene, the cropped image frame being the background model of the sub-region of the scene.
4. The method of claim 1 , wherein determining the background model of the sub-region of the scene comprises:
determining, within each of a plurality of historical image frames, one or more sub-regions being free of any foreground objects;
aggregating the one or more sub-regions from the plurality of historical image frames to form a complete background image representing the entire scene; and
cropping from the complete background image a sub-region corresponding to the sub-region of the scene, the cropped complete background image being the background model of the sub-region of the scene.
5. The method of claim 4 , wherein aggregating the one or more sub-regions from the plurality of historical image frames comprises stitching the one or more sub-regions to form an image representing the whole scene.
6. The method of claim 1 , wherein the object classifier is prepared in part using supervised learning.
7. The method of claim 1 , wherein the computer-implemented machine learning is a convolutional neural network.
8. A computer-implemented object classifier for object classification trained according to the method of claim 1 .
9. The method of claim 1 , further comprising training the object classifier by computer-implemented machine learning using a misclassified sub-region of a scene as a negative training example.
10. A system for training a computer-implemented object classifier, the system comprising:
a processor;
a computer-readable storage device storing program instructions that, when executed by the processor, cause the system to perform operations comprising:
detecting a foreground visual object within a sub-region of a scene within a field of view of a video capture device;
determining a background model of the sub-region of the scene, the background model representing the sub-region when any foreground visual object is absent therefrom;
training the object classifier by computer-implemented machine learning using the background model of the sub-region as a first negative training example, wherein the object classifier is trained specifically for a current scene;
upon the current scene being changed to a new scene, reverting to the object classifier without the training specific to the current scene; and
training the object classifier by machine learning using background models from the new scene.
11. The system of claim 10 , wherein the operations further comprise training the object classifier by machine learning using the detected foreground visual object as a positive training example.
12. The system of claim 10 , wherein determining the background model of the sub-region of the scene comprises:
selecting a historical image frame captured when any foreground object is absent from a sub-region of the historical image frame corresponding to the sub-region of the scene;
cropping from the historical image frame the sub-region corresponding to the sub-region of the scene, the cropped image frame being the background model of the sub-region of the scene.
13. The system of claim 10 , wherein determining the background model of the sub-region of the scene comprises:
determining, within each of a plurality of historical image frames, one or more sub-regions being free of any foreground objects;
aggregating the one or more sub-regions from the plurality of historical image frames to form a complete background image representing the entire scene; and
cropping from the complete background image a sub-region corresponding to the sub-region of the scene, the cropped complete background image being the background model of the sub-region of the scene.
14. The system of claim 13 , wherein aggregating the one or more sub-regions from the plurality of historical image frames comprises stitching the one or more sub-regions to form an image representing the whole scene.
15. The system of claim 10 , wherein the object classifier is prepared in part using supervised learning.
16. The system of claim 10 , wherein the computer-implemented machine learning is a convolutional neural network.
17. The system of claim 10 , wherein the operations further comprise training the object classifier by computer-implemented machine learning using a misclassified sub-region of a scene as a negative training example.