Efficient matching of robust features for embedded SLAM

Abstract

With the development of the computer technology, people try to use the camera to obtain the visual information from environment and convert it into the digital signals. The whole process of acquiring, processing, analyzing, and understanding of visual information by computer system is then developed as a new research field - Computer Vision. More and more mobile applications are equipped with camera for vision perception, environment analysis, decision making and localization. The motion of the camera system can be estimated by comparing the current frame with the previous frame. Feature based image matching approaches detect distinctive and robust features from images, find the best match between the image pair based on the similarity of features. Because of the high efficiency, robustness and noise-resistibility, image matching based on the local point features has become a widely accepted and utilized method in the recent past. A wide range of feature detectors and feature descriptors have been proposed, the performance comparison between the most known and newly proposed feature descriptors is the purpose of this thesis. A systematic comparison program is implemented to evaluate the performance of different feature descriptors under varying image deformations. The evaluation is preformed by comparing the number of keypoints, quality measures, time consumption and position error. After evaluation on the static image pairs, a real-time application is implemented for comparing the real-time performance. The results obtained in this thesis can help to choose the most suitable feature descriptors according to the application environment.