Softkinetic Software

Javier Martinez Gonzalez

Xavier Baele

Julien Thollot

Laurent Guigues

Kevin Simons

Described herein is a method of calibrating a three-dimensional imaging system. During calibration, a position and an orientation of the three-dimensional imaging system is determined with respect to a first parameter comprising a real world vertical direction (V) and to a second parameter comprising an origin of a three-dimensional scene captured by the imaging system. The first and second parameters are used to derive a calibration matrix (M) which is used to convert measurements from a virtual coordinate system (M) of the three-dimensional imaging system into a real coordinate system (M) related to the real world. The calibration matrix (M) is used to rectify measurements prior to signal processing. An inverse calibration matrix (M) is also determined. Continuous monitoring and adjustment of the setup of the three-dimensional imaging system is carried out and the calibration matrix (M) and its inverse (M) are adjusted accordingly.

Automatic scene calibration

Described herein is a method for recognizing a human head in a source image. The method comprises detecting a contour of at least part of a human body in the source image, calculating a depth of the human body in the source image. From the source image, a major radius size and a minor radius size of an ellipse corresponding to a human head at the depth is calculated, and, for at least several of a set of pixels of the detected contour, generating in an accumulator array at least one segment of an ellipse centered on the position of the contour pixel and having the major and minor radius sizes. Positions of local intensity maxima in the accumulator array are selected as corresponding to positions of the human head candidates in the source image.

Head recognition from depth image

The present invention relates to a method for tracking at least one object in a sequence of frames, each frame comprising a pixel array, wherein a depth value is associated to each pixel. The method comprises grouping at least some of said pixels of each frame into several regions, grouping said regions into clusters (B, . . . , B) of interconnected regions; and determining that a cluster (B, . . . , B) which is adjacent to another cluster (B) in a two-dimensional projection belongs to an object partially occluded by said other cluster (B) if it has a different depth value than said other cluster (B).

Growing community of inventors

Brussels, Belgium

Javier Martinez Gonzalez