A Graph-Based Vision System for Automatic Object Detection in Outdoor Scenes

Abstract

In this paper, the characteristics of the human perception system as well as the image features are exploited in a dual-resolution vision system for segmentation/object detection in outdoor scenes. The texture details are deliberately removed and similar color shades are combined together in a low-resolution version of the image, to reduce the excess image information. Using a color clustering algorithm, the color regions of the low-resolution image are found. Then, a weighted graph is constructed, whose nodes contain the detailed features of the regions, derived from the highresolution image. The weight of the edge between two nodes, Nodes' Merging Potential (NMP), denotes the advantage of merging them together to construct the fundamental image regions. This graph is then pruned regarding the NMP values, so that the main segments are developed and then identified. The proposed algorithm has shown high speed and accuracy for segmentation/object detection in outdoor scenes.