Method and apparatus for foreground object detection
View Patent ↗The present invention utilizes depth images captured by a depth camera to detect foreground/background. The method comprises establishing a single background distribution model, updating the background distribution model if a new depth value for the pixel can be represented by the background distribution model, skipping update of the background distribution model if the pixel is before the background, and replacing the background distribution model if the pixel is behind the background. In case that the background distribution model does not exist initially, a new background distribution model is created. The fluctuation of the depth value due to noise is handled by using a candidate background distribution model. Furthermore, the noise for pixels around object edges is handled by using a mixture of two background distribution models.
1. A method of background detection for a pixel around object edges from depth images using a first background distribution model or a mixture of the first background distribution model and a second background distribution model, the method comprising:
receiving a depth value for the pixel from the depth images;
updating the first background distribution model according to the depth value if the pixel can be represented by the first background distribution model;
skipping update of the first background distribution model if the pixel is before a background associated with the first background distribution model;
establishing a candidate background distribution model and evaluating a first occurrence frequency and a second occurrence frequency associated with representation of the pixel by the first background distribution model or the candidate background distribution model respectively, if the pixel is behind the background associated with the first background distribution model;
establishing the second background distribution model from the candidate background distribution model if the first occurrence frequency and the second occurrence frequency indicates high occurrence;
discarding the candidate background distribution model if the first occurrence frequency indicates the high occurrence and the second occurrence frequency indicates low occurrence;
replacing the first background distribution model by the candidate background distribution model if the first occurrence frequency indicates the low occurrence and the second occurrence frequency indicates the high occurrence; and
processing the pixel using the mixture of the first background distribution model and the second background distribution model if the second background distribution model exists, and using a single background distribution model based on the first background distribution model otherwise.
2. An apparatus for background detection for a pixel around object edges from depth images using a first background distribution model or a mixture of the first background distribution model and a second background distribution model, the apparatus comprising one or more electronic circuits configured to:
receive a depth value for the pixel from the depth images;
update the first background distribution model according to the depth value if the pixel can be represented by the first background distribution model;
skip update of the first background distribution model if the depth value indicates that the pixel is closer than a background associated with the first background distribution model;
establish a candidate background distribution model and means for evaluating a first occurrence frequency and a second occurrence frequency associated with representation of the pixel by the first background distribution model or the candidate background distribution model respectively, if the depth value indicates that the pixel is farther than the background associated with the first background distribution model;
establish the second background distribution model from the candidate background distribution model if the first occurrence frequency and the second occurrence frequency indicates high occurrence;
discard the candidate background distribution model if the first occurrence frequency indicates the high occurrence and the second occurrence frequency indicates low occurrence;
replace the first background distribution model by the candidate background distribution model if the first occurrence frequency indicates the low occurrence and the second occurrence frequency indicates the high occurrence; and
process the pixel using the mixture of the first background distribution model and the second background distribution model if the second background distribution model exists, and using a single background distribution model based on the first background distribution model otherwise.