earticle

영어

This paper presents a non-stationary object detection method by exploring time-varying spatial domain information in full motion video. Initially, the edge maps of difference image between two adjacent frames and current frame is generated via the well-known Canny edge detector. The distance of the edge pixels between the difference image and the current video frame are confined within a small value to determine the initial edge mask for the object in motion. The horizontal and vertical filling followed by morphological opening and closing operator are applied on the initial edge mask to create initial temporal segmentation mask of the moving object. The morphological dilation and corrosion operator are utilized to obtain binary marker image of the foreground and background which are used to modify the multi-scale morphological gradient image of current frame. Finally, the watershed algorithm is performed on the modified gradients to find the non-stationary objects accurately in the spatial domain of motion frames. Processed video results show detection accuracy of 98% and 99% for four different video experimentation test-beds involving fast and slow human motion. In this operation, the proposed technique eliminates the problem of over-segmentation of the watershed algorithm and extracts visually distinct, contextually meaningful non-stationary objects as they randomly appear (or disappear) in video sequences.