Design and Implementation of a Smart-phone Based Location System

• Main Authors: Lo, Yu-Ching, 羅郁晴
• Other Authors: Yeh, Chun-Chao
• Format: Others
• Language: zh-TW
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/11326655829912832052

碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 102 === When people in a strange place, they might want to know where they are and the surrounding location information. The purpose of this paper is to design a smart-phone based positioning system. The system allows users to get their current location information via their smart-phone. The client software for user smart-phone can be distributed and installed easily through standard mobile App. Once the client App is ready in users' smart-phone, they just need to take a picture at target objects arround, and upload the picture to the positioning system. The system will identify the location, based on the upload picture and other related information. To evaluate system performance of the proposed mechanism, intensive experiments had been done. we evaluated classification performance with different classifiers, and the impact of traning image samples and the resolution quality of the training/testing images. The images for the experiments are from a outdoor scenario (in a campus). We collected more than 20 thousands of images took from the outdoor scenario at different days. The images were classified into 17 classes, based on the location the images belonging to. From the experiment results, we found that under well control of image quality for both training and testing images the correct classification rate can be as high as 99.9%. For real user test, the correct rate drops to 69.5%. Comparing with the correct rate via random guest (1/17=5.88%), it is still prety good. Two main reasons contribute to the in-accuracy. First, many of the user testing images are focused at different spots and with different vision to some degrees, comparing with the images in the training images. we found many of the testing images are with four times wide vision than the training images. Second, image resolution is much different. All the training images were took with a resolution of 640x480 pixels, while all the testing images were took with a resolution of 3920x2204 pixels. We expect the correct rate for free users can be improved by increasing image resolution and adding different images took under different visions and DOFs (depth of fields).