Salient region detection based on stereo vision

Traditional pixel-based saliency model has some deficiencies, such as poorly defined borders and low performance in low contrast situation. A stereo vision based salient region detection approach was proposed. Simple linear iterative clustering (SLIC) method was adopted to perform superpixel segmentation. Superpixels were merged to construct segmentation image. Depth cue was computed by measuring the distance of region shifts in given stereo pair. For each region, color contrast and depth contrast were computed separately, and then fused to get saliency value. Experimental result shows that saliency map has clear contour and sharp edge and regions at close range or with high depth contrast get more saliency. The proposed method is consistent with human visual perception and suitable for obstacle detection and scene recognition in mobile robot.