Defesa de Exame de Qualificação – Alexandre Reeberg de Mello – 22/06/2017
22/06/2017 16:57
| Defesa de Exame de Qualificação | |
| Aluno | Alexandre Reeberg de Mello |
| Orientado | Prof. Marcelo Ricardo Stemmer, Dr. – DAS/UFSC |
| Data
Local |
22/06/2017 08h30 (quinta-feira)
Sala PPGEAS I (piso superior) |
| Prof. Ubijrajara Franco Moreno, Dr. – DAS/UFSC (presidente)
Prof. Aldo Von Wangenheim, Dr. – INE/UFSC Prof. Adilson Gonzaga, Dr. – DEE/USP |
|
| Título
|
A Framework for Online Object Recognition in an RGB-D Environment for Mobile Robot Object Search Task |
| Abstract: Recognition is a core problem of learning visual categories, and has a variety of potential applications that meets many areas of artificial intelligence and information retrieval, for example, object identification for mobile robots. Object recognition is an essential ability for building object-based representations of the environment or manipulate objects, and to recognize objects under an uncontrolled real-world environment is a challenging task. In this manner, this thesis proposal comes up with a framework to recognize objects during a robot movement, considering that the environment is represented by RGB-D images, and the recognition is made by a processing pipeline that includes: image processing, object features extraction, and online learning method. The RGB-D image acquisition method is chosen because it integrates view-based information from RGB and a depth information, which may turn the object recognition system more robust. A saliency detection method must be applied to identify potential objects in scene, therefore several methods will be studied and tested, and the same research procedure is for chose the segmentation algorithm, which separates the object from background and foreground. Objects and scenes data are divided in two groups: training data (that is composed of a controlled environment dataset), and test data, that is acquired by a Microsoft Kinect sensor in a real environment. Each object is represented by a set of descriptors that summarize the object features into high-dimensional arrays, and as each type of descriptor extracts a particular object characteristic, different types of descriptors will be tested. The learning and classifier system is an online Support Vector Machine (SVM), which is fed by large data of high-dimensional object descriptors. Thus, the recognition system must have relatively fast training method, to be able to be online updated, i.e., retrain the SVM model to update its support vectors. The SVM formulation includes a non-linear kernel, that must be chosen or designed by an optimized kernel hyperparameters strategy. The classifier should separate the objects in many classes, thus a multiclass strategy for the SVM will be studied and proposed. For last, a strategy to check if the classifier output is consistent with time sequenced data is introduced. | |
