University of Twente Student Theses


Recognition of objects in RGB-D data of building interior

Li, Kezhen (2013) Recognition of objects in RGB-D data of building interior.

[img] PDF
Abstract:With more and more demand for accurate indoor model, accurate 3d indoor reconstruction and recognition is needed. In order to build the indoor models, the first procedure is to recognize walls, floors, ceilings and objects of indoor environments in the acquired data. Acquired by the emerging new sensor Kinect, the RGB-D data can be divided into RGB color image and depth image .Our research aims at taking advantage of both RGB and depth to develop an effective and robust methodology to recognize indoor objects using the RGB-D data acquired by Kinect. In this research, a feature based recognition and classifier training method are combined. First step, gradient descriptor and size descriptor are adopted to extract corresponding local features from RGB, depth and point clouds of the training data. A large dataset with multi-view of objects are used as the training data. Those local features are then quantized and clustered using the bag of words concept. A histogram which summarizes the local features of each training image is later generated. The visual words of the cluster are made for further use. After then, using those features over the training data, the SVM classifier and QDC classifier are trained for discriminating different classes of objects. While on the other hand, we apply the same feature extraction using the visual words made previously to our test image. The recognition is conducted by applying the trained classifier to the test features. Two recognition experiments are carried out in our work: 1. self test recognition. 2. Recognition of data of real scene which contain multiple objects and background. Conclusion is drawn from the experiment and the recognition result that using SVM as the training classifier, the combination of the features from RGB and depth data has a superior performance over the individual one. Keywords: object recognition, RGB-D data, gradient feature, size feature, bag of words, SVM classifier
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page