Scalable object classification using range images

2011-12-13Object classification using depth images has been actively studied in robotics and computer vision fields to autonomously recognize 3D objects in the scene even under no or bad illumination based on surface geometry. The traditional protocol of object classification is to manually collect and annotate training image examples and then learn models to explain the annotated examples. However, when a large number of images are given (e.g. content based image retrieval) and/or a system has to learn a sequence of new images into the existing model on the fly (e.g. autonomous robotic exploration), this approach requires a huge amount of human effort in image annotation. ❧ In this thesis, I present an automatic and scalable framework for object classification using range images, which has been rarely discussed before. The proposed system is designed to 1) build a hierarchical structured database from a collection of unlabeled range images, 2) classify a new image, i.e. the partial surface of an 3D object, into one of existing classes or a new class using the hierarchical model of object classes and 3) automatically update the existing model of object classes to insert the new object when its surface characteristic is not well trained in the current model. Each path in the tree corresponds to a group of range images with similar surface characteristics. The class inference and online learning processes can be performed efficiently due to the tree structure of the model. ❧ Another contribution of my research is to utilize two crucial context cues, 1) 3D spatial constraints between objects and large structures and 2) the visibility context, for perception of of 3D objects. The proposed framework has been extensively validated on a large synthetic range image dataset and real depth images acquired from real-time range sensors. ❧ In this thesis, I also present two methods to tackle the problem of object segmentation in range images. The first method segments the partial surface of each object from 3D aerial scans of a vast urban area captured from LIDAR sensors. It enforces the spatial context between large structures and objects. The second work employs prior information of objects in order to deal with cluttered environments. This thesis finally delineates a method of 3D object recognition and pose estimation using the visibility context and demonstrates that the proposed approach outperforms state-of-the-art methods