Pixel-level based micro-feature extraction
View Patent ↗Techniques are disclosed for extracting micro-features at a pixel-level based on characteristics of one or more images. Importantly, the extraction is unsupervised, i.e., performed independent of any training data that defines particular objects, allowing a behavior-recognition system to forgo a training phase and for object classification to proceed without being constrained by specific object definitions. A micro-feature extractor that does not require training data is adaptive and self-trains while performing the extraction. The extracted micro-features are represented as a micro-feature vector that may be input to a micro-classifier which groups objects into object type clusters based on the micro-feature vectors.
1. A computer-implemented method for extracting pixel-level micro-features from image data captured by a video camera, the method comprising:
receiving the image data;
identifying a set of pixels in the image data associated with a foreground patch that depicts a foreground object;
evaluating appearance values of the pixels included in the set of pixels to compute a plurality of micro-feature values representing the foreground object, each based on at least one pixel-level characteristic of the foreground patch, wherein the micro-feature values are computed independent of training data that defines a plurality of object types;
generating a micro-feature vector that includes the plurality of micro-feature values;
classifying the foreground object as depicting an object type as based on the micro-feature vector, wherein the object type is determined by mapping the micro-feature vector to a cluster in a self-organizing map (SOM) adaptive resonance theory (ART) network generated from a plurality of micro-feature vectors; and
updating one or more cluster properties associated with the cluster based on the plurality of micro-feature values in the generated micro-feature vector.
2. The computer-implemented method of claim 1 , wherein a bounding box minimally encloses the foreground patch and one of the computed micro-feature values is a normalized pixel-level area of the foreground patch.
3. The computer-implemented method of claim 1 , wherein a bounding box minimally encloses the foreground patch and one of the computed micro-feature values is an aspect ratio of the bounding box.
4. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is set to predetermined value indicating that the computed micro-feature value is not valid when a threshold value is not met.
5. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is a hue entropy value and the pixel-level characteristic is hue values of the foreground patch.
6. The computer-implemented method of claim 1 , one of the computed micro-feature values is a normalized magnitude-saturation ratio value and the pixel-level characteristic is magnitude and saturation values of the foreground patch.
7. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is an absolute value of a cosine of an angle of alignment between a major axis and an orientation line of the foreground patch and the pixel-level characteristic is the orientation line of the foreground patch.
8. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is a groupiness value and the pixel-level characteristic is a horizontal distribution and vertical distribution of the foreground patch.
9. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is a legged-ness value based on a sum of angles between neighboring segments of a start skeleton of the foreground patch and the pixel-level characteristic is the star skeleton of the foreground patch.
10. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is a verticality value and the pixel-level characteristic is instantaneous per-pixel gradients of the foreground patch.
11. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is an animateness value and the pixel-level characteristic is a covariance matrix of the foreground patch in a first frame and a second frame of the image data.
12. The computer-implemented method of claim 1 , wherein one of the computed micro-feature values is a periodicity value and the pixel-level characteristic is an auto-correlation of a self-similarity matrix computed based on a sequence of frames including the foreground patch.
13. A non-transitory computer-readable storage medium containing a program which, when executed by a processor, performs an operation for extracting pixel-level micro-features from image data captured by a video camera, the operation comprising:
receiving the image data;
identifying a set of pixels in the image data associated with a foreground patch that depicts a foreground object;
evaluating appearance values of the pixels included in the set of pixels to compute a plurality of micro-feature values representing the foreground object, each based on at least one pixel-level characteristic of the foreground patch, wherein the micro-feature values are computed independent of training data that defines a plurality of object types;
generating a micro-feature vector that includes the micro-feature value; and
classifying the foreground object as depicting an object type as based on the micro-feature vector, wherein the object type is determined by mapping the micro-feature vector to a cluster in a self-organizing map (SOM) adaptive resonance theory (ART) network generated from a plurality of micro-feature vectors; and
updating one or more cluster properties associated with the cluster based on the plurality of micro-feature values in the generated micro-feature vector.
14. The computer-readable storage medium of claim 13 , wherein one of the computed micro-feature value is a hue entropy value and the pixel-level characteristic is hue values of the foreground patch.
15. The computer-readable storage medium of claim 13 , wherein one of the computed micro-feature value is a normalized magnitude-saturation ratio value and the pixel-level characteristic is magnitude and saturation values of the foreground patch.
16. The computer-readable storage medium of claim 13 , wherein one of the computed micro-feature value is an absolute value of a cosine of an angle of alignment between a major axis and an orientation line of the foreground patch and the pixel-level characteristic is the orientation line of the foreground patch.
17. The computer-readable storage medium of claim 13 , wherein one of the computed micro-feature value is a groupiness value and the pixel-level characteristic is a horizontal distribution and vertical distribution of the foreground patch.
18. A system, comprising:
a video input source configured to provide image data;
a processor; and
a memory containing a program, which, when executed on the processor is configured to perform an operation for extracting pixel-level micro-features from the image data captured by the video input source, the operation comprising:
receiving the image data;
identifying a set of pixels in the image data associated with a foreground patch that depicts a foreground object;
evaluating appearance values of the pixels included in the set of pixels to compute a plurality of micro-feature values representing the foreground object, each based on at least one pixel-level characteristic of the foreground patch, wherein the micro-feature values are computed independent of training data that defines a plurality of object types;
generating a micro-feature vector that includes the micro-feature value;
classifying the foreground object as depicting an object type as based on the micro-feature vector, wherein the object type is determined by mapping the micro-feature vector to a cluster in a self-organizing map (SOM) adaptive resonance theory (ART) network generated from a plurality of micro-feature vectors; and
updating one or more cluster properties associated with the cluster based on the plurality of micro-feature values in the generated micro-feature vector.
19. The system of claim 18 , wherein a bounding box minimally encloses the foreground patch and one of the computed micro-feature values is a normalized pixel-level area of the foreground patch or an aspect ratio of the bounding box.
20. The system of claim 18 , wherein one of the computed micro-feature values is an absolute value of a cosine of an angle of alignment between a major axis and an orientation line of the foreground patch and the pixel-level characteristic is the orientation line of the foreground patch.
21. The system of claim 18 , wherein one of the computed micro-feature values is a groupiness value and the pixel-level characteristic is a horizontal distribution and vertical distribution of the foreground patch.
22. The system of claim 18 , wherein one of the computed micro-feature values is a legged-ness value based on a sum of angles between neighboring segments of a start skeleton of the foreground patch and the pixel-level characteristic is the star skeleton of the foreground patch.
23. The system of claim 18 , wherein one of the computed micro-feature values is a verticality value and the pixel-level characteristic is instantaneous per-pixel gradients of the foreground patch.
24. The system of claim 18 , wherein one of the computed micro-feature values is an animateness value and the pixel-level characteristic is a covariance matrix of the foreground patch in a first frame and a second frame of the image data.
25. The system of claim 18 , wherein one of the computed micro-feature values is a periodicity value and the pixel-level characteristic is an auto-correlation of a self-similarity matrix computed based on a sequence of frames including the foreground patch.