Neural model-based segmentation of image motion

Besides enabling the segmentation of video streams into moving and background components, detecting moving objects provides a focus, of attention for recognition, classification, and activity analysis, making these later steps more efficient. We propose a novel model for image sequences based on self organization through artificial neural networks, that is used both for background modeling, allowing to handle scenes containing moving backgrounds or gradual illumination variations, and for stopped foreground modeling, helping ill distinguishing between moving and stopped foreground regions and leading to an initial segmentation of scene objects. Experimental results are presented for real video sequences.