Stacked generalization for scene analysis and object recognition

The problem of object recognition and detection has been largely addressed by the robotics community, since its importance both in mapping and manipulation problems. One possible approach for the recognition task is to assume a specific a-priori knowledge of the objects possibly present in a scene. In this framework, this paper presents a novel technique for object detection and recognition based on Stacked Generalization (SG) method developed by Wolpert in 1992. The innovation of the proposed technique is the introduction of SG classification method to perform a multi-layer object recognition fusing heterogeneous spatial and color data acquired with an RGB-D camera. To improve the accuracy and the robustness of the system to environmental variability, we introduce a second layer classifier. Its goal is to evaluate and weights the results of the first layer classifiers, thus combining and improving the overall classification performance. This technique has a low computational cost and is suitable for on-line applications, such as robotic manipulation or automated logistic systems. To validate the presented approach experimental tests have been carried out and results are reported.