OBJECT EXTRACTION AND DEPTH RECOVERY FROM STEREO IMAGE PAIRS USING A NEURAL NETWORK BASED PIXELCLASSIFICATION METHOD

摘要

The present research deals with depth recovery by object extraction and tracking in stereo-image pairs. A Self Organizing Multi Layer Perceptron is developed to extract object masks from the left image by classifying foreground and background pixels. Subsequently, K-means clustering is employed on the foreground pixels to find the number of objects in the scene. For each object, a modified Block-matching algorithm is applied on the corresponding right image to calculate the disparity. The depth of the object is then derived from the disparity by using triangulation. The combination of three methods - Self Organizing MLP, K-means, and Block Matching for depth recovery is useful for automating the process of multi object extraction. This approach can be applied to 3D reconstruction, autonomous guidance, robotic vision, and virtual reality.