2d/3d Object Recognition And Categorization Approaches For Robotic Grasping

Object categorization and manipulation are critical tasks for a robot to operate in the household environment. In this chapter, we propose new methods for visual recognition and categorization. We describe 2D object database and 3D point clouds with 2D/3D local descriptors which we quantify with the k-means clustering algorithm for obtaining the bag ofwords (BOW). Moreover, we develop a new global descriptor called VFH-Color that combines the original version of Viewpoint Feature Histogram (VFH) descriptor with the color quantization histogram, thus adding the appearance information that improves the recognition rate. The acquired 2D and 3D features are used for training Deep Belief Network (DBN) classifier. Results from our experiments for object recognition and categorization show an average of recognition rate between 91% and 99% which makes it very suitable for robot-assisted tasks.