NAVIGATION AND MAPPING THROUGH AUTOMATED IMAGE UNDERSTANDING AND RETRIEVAL

Open Access
- Author:
- Valluri, Adityagiri
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- April 01, 2009
- Committee Members:
- David Jonathan Miller, Thesis Advisor/Co-Advisor
David Jonathan Miller, Thesis Advisor/Co-Advisor
James Z Wang, Thesis Advisor/Co-Advisor - Keywords:
- image retrieval
mapping
navteq database
SIFT - Abstract:
- A novel method for image retrieval of navigational data through automated image understanding is presented, for mapping purposes. The method discussed here involves two steps, an (a) Offine process and an (b) Online process. In the offine process, the database of images would be clustered roughly into groups based on their color histograms using the LSH algorithm. Further, the images are processed using the graph based visual saliency method to highlight the regions with color and as a result a reduction of 62.3% is achieved in the number of SIFT features in the next step. SIFT features are extracted from these resulting images. In the online process, a query image is provided and is assigned to one of the groups created in the offline process based on a distance measure. We employ the use of SIFT features, opponent color space in feature extraction. In feature matching, we make use of histogram intersection score filter and a robust matching procedure proposed by Lowe. We compare our methods to many state of the art algorithms in logo detection and object recognition. Our method does not assume apriori knowledge of the candidate logos in the database. We state that the image database is logo-rich and we set out to find any image that is presented to the retrieval system. Experimental results confirm that our methods perform at least 25% better than Lowe's object recognition algorithm and that our methods are robust to changes in scale, illumination and angle of rotation.