An Object-driven Online Segmentation System for Mobile Robots

In this paper we propose an object-driven online segmentation system for mobile robots. Among existing tracking and segmentation methods, the methods themselves are highly emphasized while properties of objects are usually not exploited enough. We propose an adaptive selection mechanism based on the properties of the objects to automatically choose an optimal tracking algorithm. For textured objects, a tracking algorithm which combines Lucas-Kanade tracker and model based detector using Random Forests classifier is adopted; for uniform objects, a color based tracker with a smooth constraint enforcement is used to ensure a robust performance. Moreover, a two-step object segmentation using Gaussian Mixture Models and graph cuts is applied to obtain detailed shape information. The experimental results on a variety of objects show the effectiveness of the adaptive tracking selection mechanism. The system also yields very promising performance in very challenging conditions with occlusions, illumination changes as well as cluttered background.