OBJECT DETECTION AND EXTRACTION FROM IMAGE SEQUENCES

摘要

Object detection and extraction is performed from image sequences utilizing circular buffers for both source images and tracking. The detection and extraction process is performed in relation to the previous and current image, and the current and next image, including: alignment, absolute difference, removal of non-overlaps, and contour detection in difference images. An intersection is performed on these two outputs to retain contours of the current image only. Recovery of missing contour information is performed utilizing gradient tracing, followed by morphological dilation. A splitting process is performed if additional objects are found in a bounding box area. A mask image bounded by object contour is created, color attributes assigned, object verification performed and outliers removed. Then untracked objects are removed from the mask and a mask is output for moving objects with rectangular boundary box information.

主权项

1. An apparatus for performing object detection and extraction from image sequences, comprising:
(a) a computer processor; and (b) programming in a non-transitory computer readable medium and executable on said computer processor for performing steps comprising:
(i) downsizing images from an image sequence;(ii) buffering at least three sequential images of the image sequence in a circular buffer;(iii) aligning a previous image and a next image to a current image of said image sequence and outputting a previous difference between the previous image and the current image, and a next difference between the current image and the next image;(iv) determining absolute differences between an aligned previous image and an aligned current image, and between the aligned current image and an aligned next image;(v) removing non-overlaps which are redundant non-corresponding sections due to the non-overlap;(vi) detecting contours on each difference image;(vii) performing an intersection between the previous difference and the next difference to retain contours of the current image only;(viii) performing gradient tracing to recover any missing object contour information;(ix) generating an object mask and assigning color attributes to a bounding box; and(x) eliminating outliers and clearing the object mask of untracked objects to output a binary mask configured for moving objects and rectangular bounding box information;whereby said apparatus automatically extracts dynamic moving objects from a sequence of images captured from cameras which do not require static platforms.