Abstract: A system and method of generating a composite frame from a video capturing a plurality of persons. The method comprises determining a plurality of frames corresponding to occurrence of a predetermined event in the video; determining an importance score for each of the plurality of persons in the determined plurality of frames, the importance score determined based on a type of interaction detected between each of the plurality of persons with an activity object; and selecting one or more of the plurality of persons based on the determined importance score to generate a composite frame of the event including the selected one or more persons.

Abstract: A method of detecting background pixels in a video, the video including a sequence of frames, each frame having a pitch and people. The method includes segmenting the pitch and determining a height of one of the people. An inclusion mask for the sequence of frames is then created using the pitch segmentation and the people's height. Each of the sequence of frames is segmented into foreground and background pixels. The method then creates a foreground appearance model from the foreground pixels inside said inclusion mask over the sequence of frames, and a background appearance model from the background pixels inside said inclusion mask over the sequence of frames. The method then uses the created foreground and background appearance models to segment one of the sequence of frames into definite foreground, definite background and uncertain pixels.

Abstract: A method of detecting background pixels in a video, the video including a sequence of frames, each frame having a pitch and people. The method includes segmenting the pitch and determining a height of one of the people. An inclusion mask for the sequence of frames is then created using the pitch segmentation and the people's height. Each of the sequence of frames is segmented into foreground and background pixels. The method then creates a foreground appearance model from the foreground pixels inside said inclusion mask over the sequence of frames, and a background appearance model from the background pixels inside said inclusion mask over the sequence of frames. The method then uses the created foreground and background appearance models to segment one of the sequence of frames into definite foreground, definite background and uncertain pixels.

Abstract: A system and method of generating a composite frame from a video capturing a plurality of persons. The method comprises determining a plurality of frames corresponding to occurrence of a predetermined event in the video; determining an importance score for each of the plurality of persons in the determined plurality of frames, the importance score determined based on a type of interaction detected between each of the plurality of persons with an activity object; and selecting one or more of the plurality of persons based on the determined importance score to generate a composite frame of the event including the selected one or more persons.

Abstract: A method of determining a vertical medial axis of an object in an input image, is disclosed. A measure of horizontal symmetry is determined for each pixel of the image. First and second vertical bounds within the image are determined. A continuous path between the first and second vertical bounds is determined based on the horizontal symmetry measure, the continuous path representing the vertical medial axis of the object.

Abstract: A method of modifying a scene model for a sequence of input images includes receiving, selecting, determining, and modifying. Scene modes associated with the sequence are received. For each image pixel in an input image of the sequence, two scene modes are selected from the received scene modes based on distances between the image pixel color value and the scene modes color values at a corresponding pixel location. It is determined if the image pixel color value matches the color value of any of the selected scene modes. It is determined if the image pixel color value is a mixture of the color values of the selected scene modes. The color value of a closest selected scene mode is modified or is maintained depending on whether there is one matched mode and the image pixel color value is not a mixture of the color values of the selected scene modes.

Abstract: A method of classifying foreground and background in an image of a video by determining a pitch color model of the image, the pitch color model comprising a pitch color, color shades of the pitch color, and the pitch color and the color shades under different shades of shadow. Then determining a pitch mask based on a pitch segmentation of the image of the video, determining a pitch background model based on the pitch mask and the pitch color model. The method may continue by classifying each of the elements of the pitch mask as background if a color of the element of the pitch mask matches the pitch color model and updating the pitch background model and the pitch color model using the colors of the elements that have been classified to match the pitch color model.

Abstract: A method of classifying foreground and background in an image of a video by determining a pitch colour model of the image, the pitch colour model comprising a pitch colour, colour shades of the pitch colour, and the pitch colour and the colour shades under different shades of shadow. Then determining a pitch mask based on a pitch segmentation of the image of the video, determining a pitch background model based on the pitch mask and the pitch colour model. The method may continue by classifying each of the elements of the pitch mask as background if a colour of the element of the pitch mask matches the pitch colour model and updating the pitch background model and the pitch colour model using the colours of the elements that have been classified to match the pitch colour model.

Abstract: Method, system, apparatus and computer readable medium for re-identifying a query objection moving from a first camera view to a second camera view. The method comprises the steps of accessing a first set of image frames (Q) captured from the first camera view and a second set of image frames (C) captured from the second camera view; selecting a first representative frame (QR) from the first set of image frames and a second representative frame (CR) from the second set of image frames based on a plurality of intra-camera object appearance distances between frame pairs in each set of image frames; and re-identifying the candidate object in the second camera view as the query object in the first camera view based on at least one inter-camera object appearance distance (D) calculated between the first representative frame and the second representative frame.

Abstract: A method of modifying a scene model for a sequence of input images includes receiving, selecting, determining, and modifying. Scene modes associated with the sequence are received. For each image pixel in an input image of the sequence, two scene modes are selected from the received scene modes based on distances between the image pixel color value and the scene modes color values at a corresponding pixel location. It is determined if the image pixel color value matches the color value of any of the selected scene modes. It is determined if the image pixel color value is a mixture of the color values of the selected scene modes. The color value of a closest selected scene mode is modified or is maintained depending on whether there is one matched mode and the image pixel color value is not a mixture of the color values of the selected scene modes.

Abstract: A method and system associated with a camera view of a moving-object in a scene. The method comprises detecting and tracking the moving object over multiple video frames, estimating an orientation of the moving object in each of the video frames, and constructing a cost map from the estimated orientations over the multiple video frames for finding a minimum cost path over the cost map. The Method also comprises determining regularized orientation estimates of the moving-object from the minimum cost path, and locating the vanishing point of the camera view based on an axis of the moving-object from the minimum cost path, the axis formed by using the regularized orientation estimates.

Abstract: A method of determining stability of a camera comprises capturing an image with the camera and then (A) determining a magnitude of motion for each region of a current sub-division of the image. A step (B) then determines a number of the magnitudes of motion not larger than a magnitude threshold associated with the current sub-division, and where the determined number is greater than or equal to a region threshold associated with the number of regions in the current sub-division, (i) determining the camera to be stable, or otherwise (ii) dividing the current sub-division of the image into a further sub-division and repeating steps (A) and (B) upon the further sub-division.

Abstract: A method of updating a scene model for a foreground segmentation of an input image captured from a camera, is disclosed. One or more visual elements of the input image are determined. A spatial relationship between at least one of the visual elements and the scene model for a foreground segmentation of the input image is determined. The method updates the scene model for determining the foreground segmentation of the input image based on the determined spatial relationship.

Abstract: At least one method of, and at least one apparatus, system and computer readable medium for, matching moving targets between two camera views are discussed herein. The moving targets are tracked over video frames in each of the camera views and motion directions of the targets are recorded. Appearance similarities for cross-camera target pairs over synchronized ones of the video frames are determined. A joint probability model of co-occurred ones of the recorded motion directions for the camera views are determined using a weighted accumulation of the recorded motion directions of the cross-camera target pairs, where accumulation weights used in determining the joint probability model are derived from the determined appearance similarities. The moving targets between the camera views are matched using the joint probability model.

Abstract: A method and system associated with a camera view of a moving-object in a scene. The method comprises detecting and tracking the moving object over multiple video frames, estimating an orientation of the moving object in each of the video frames, and constructing a cost map from the estimated orientations over the multiple video frames for finding a minimum cost path over the cost map. The Method also comprises determining regularised orientation estimates of the moving-object from the minimum cost path, and locating the vanishing point of the camera view based on an axis of the moving-object from the minimum cost path, the axis formed by using the regularised orientation estimates.

Abstract: At least one method of, and at least one apparatus, system and computer readable medium for, matching moving targets between two camera views are discussed herein. The moving targets are tracked over video frames in each of the camera views and motion directions of the targets are recorded. Appearance similarities for cross-camera target pairs over synchronised ones of the video frames are determined. A joint probability model of co-occurred ones of the recorded motion directions for the camera views are determined using a weighted accumulation of the recorded motion directions of the cross-camera target pairs, where accumulation weights used in determining the joint probability model are derived from the determined appearance similarities. The moving targets between the camera views are matched using the joint probability model.

Abstract: Method, system, apparatus and computer readable medium for re-identifying a query objection moving from a first camera view to a second camera view. The method comprises the steps of accessing a first set of image frames (Q) captured from the first camera view and a second set of image frames (C) captured from the second camera view; selecting a first representative frame (QR) from the first set of image frames and a second representative frame (CR) from the second set of image frames based on a plurality of intra-camera object appearance distances between frame pairs in each set of image frames; and re-identifying the candidate object in the second camera view as the query object in the first camera view based on at least one inter-camera object appearance distance (D) calculated between the first representative frame and the second representative frame.

Abstract: A method of blurring an image using a spatially variable blur is disclosed. The method comprises the steps of: for each pixel of a scan line of pixels in the image: obtaining a measure representative of image depth at the pixel; determining a blur kernel width for the pixel based on the measure representative of image depth at the pixel; and where an increase in the determined blur kernel width with respect to a blur kernel width of a previous pixel in the scan line is detected: determining, from the determined blur kernel width, a tapered blur width adapted to provide an asymmetric filter response for the pixel; and applying the tapered blur width to a filter to output a blurred pixel value for the pixel by using a plurality of filter coefficients determined based on blur width values of one or more previously blurred pixels.

Abstract: A method of determining a vertical medial axis of an object in an input image, is disclosed. A measure of horizontal symmetry is determined for each pixel of the image. First and second vertical bounds within the image are determined. A continuous path between the first and second vertical bounds is determined based on the horizontal symmetry measure, the continuous path representing the vertical medial axis of the object.

Abstract: A method of producing an image with a blurred background (151) from an initial image (111) and a depth map (112) identifying a foreground region (236) and a background region (237) of the image, said method comprising determining (670) an edge visibility measure (671) for one edge (231) in the depth map (112) by comparing (820) a pixel value (233) in the foreground with a pixel value (234) in the background, said edge visibility measure (671) distinguishing between the foreground (236) and the background (237); applying (680) a spatially variable blur to the depth map (112) to produce a compositing matte (681), dependent on the edge visibility measure (671); applying (530) a blur to the background region (521) of the depth map (112); and compositing (150) the blurred background image data (531) with the initial image (111) to produce the rendered image with the blurred background (151).