Identification of people using multiple types of input
AssigneeMicrosoft Corporation
InventorsCha Zhang, Paul A. Viola, Pei Yin, Ross G. Cutler, Xinding Sun, Yong Rui
Life of this patent
Estimating the term…
Abstract
Systems and methods for detecting people or speakers in an automated fashion are disclosed. A pool of features including more than one type of input (like audio input and video input) may be identified and used with a learning algorithm to generate a classifier that identifies people or speakers. The resulting classifier may be evaluated to detect people or speakers.
Claims18
A method, comprising:
- identifying a pool of features from multiple types of input, the pool of features comprising a first video feature from a video input, the first video feature comprising a parent rectangle within a feature rectangle in an image of the video input;
- calculating a numeric value associated with the first video feature by summing values of pixels in the parent rectangle;
- generating a classifier for speaker detection using a learning algorithm wherein nodes of the classifier are selected using the pool of features, based on the numeric value associated with the first video feature;
- dividing an image of a second video feature into a plurality of detection windows;
- evaluating the classifier against input data for each detection window to produce an estimate of likelihood that a person exists in the detection window;
- and
- detecting a person in a region in the image by selecting a region that includes a relatively high number of detection windows that each have a high estimate of likelihood of containing a person.
An apparatus, comprising a detector device including a detector configured to accept multiple types of input, the multiple types of input comprising a video input, the detector configured to evaluate a person detection classifier to detect a person, the person detection classifier created by:
- identifying a pool of features from multiple types of input, the pool of features comprising a first video feature from a video input, the first video feature comprising a parent rectangle within a feature rectangle in an image of the video input;
- calculating a numeric value associated with the first video feature by summing values of pixels in the parent rectangle;
- generating the classifier for person detection using a learning algorithm wherein nodes of the classifier are selected using the pool of features, based on the numeric value associated with the first video feature;
- dividing an image of a second video feature into a plurality of detection windows;
- evaluating the classifier against input data for each detection window to produce an estimate of likelihood that a person exists in the detection window;
- and
- detecting a person in a region in the image by selecting a region that includes a relatively high number of detection windows that each have a high estimate of likelihood of containing a person.
A computer-readable storage medium containing instructions that when executed cause one or more processors to:
- identify a pool of features from multiple types of input, the pool of features comprising a first video feature from a video input, the first video feature comprising a parent rectangle within a feature rectangle in an image of the video input;
- calculate a numeric value associated with the first video feature by summing values of pixels in the parent rectangle;
- generate a classifier for speaker detection using a learning algorithm wherein nodes of the classifier are selected using the pool of features, based on the numeric value associated with the first video feature;
- divide an image of a second video feature into a plurality of detection windows;
- evaluate the classifier against input data for each detection window to produce an estimate of likelihood that a person exists in the detection window;
- and
- detect a person in a region in the image by selecting a region that includes a relatively high number of detection windows that each have a high estimate of likelihood of containing a person.
Dependencies are inferred from the claim text.