In this paper we present a novel vision system for object-driven and online learning based segmentation of unknown objects in a scene. The main application of this system is for mobile robots exploring unknown environments, where unknown objects need to be inspected and segmented from multiple viewpoints. In an initial step, objects are detected using a bottom-up segmentation method based on salient information. The cluster with the most salient points is assumed to be the most dominant object in the scene and serves as an initial model for online segmentation. Then the dominant object is tracked by a Lucas-Kanade tracker and the object model is constantly updated and learned online based on Random Forests classifier. To refine the model a two-step object segmentation using Gaussian Mixture Models and graph cuts is applied. As a result, the detailed contour information of the dominant unknown object is obtained and can further be used for object grasping and recognition. We tested our system in very challenging conditions with multiple identical objects, severe occlusions, illumination changes and cluttered background and acquired very promising results. In comparison with other methods, our system works online and requires no input from users.
Paper
Full text
Robust Online Segmentation of Unknown Objects for Mobile Robots
Semantic Scholar · Computer Science · 2012
Abstract
In this paper we present a novel vision system for object-driven and online learning based segmentation of unknown objects in a scene. The main application of this system is for mobile robots exploring unknown environments, where unknown objects need to be inspected and segmented from multiple viewpoints. In an initial step, objects are detected using a bottom-up segmentation method based on salient information. The cluster with the most salient points is assumed to be the most dominant object in the scene and serves as an initial model for online segmentation. Then the dominant object is tracked by a Lucas-Kanade tracker and the object model is constantly updated and learned online based on Random Forests classifier. To refine the model a two-step object segmentation using Gaussian Mixture Models and graph cuts is applied. As a result, the detailed contour information of the dominant unknown object is obtained and can further be used for object grasping and recognition. We tested our system in very challenging conditions with multiple identical objects, severe occlusions, illumination changes and cluttered background and acquired very promising results. In comparison with other methods, our system works online and requires no input from users.