| Summary: | 碩士 === 南台科技大學 === 資訊工程系 === 102 === We propose a new concept of structural sampling in this thesis for two computer vision applications, pupil localization and lane line detection, wherein object structures of pupil and of lane lines are mathematically described and incorporated into the sampling processing of image features/blocks for the detection of target objects. In pupil localization, the sampling of edge points based on contour structures is employed to allocate pupil circles in an image sequence. For lane line detection, new block samples of current image frame are generated based on the lane line structure estimated in previous image frame.
In the first application, a new pupil localization system is developed for excimer laser eye surgery to locate a proper pupil center prior to the laser surgery. Our system design includes four processing steps: (1) adaptive edge extraction, (2) specular region reduction, (3) contour-based sampling, and (4) least-square fitting. Particularly, in the third processing step, we propose a novel circle detection method using structural sampling and demonstrate its effectiveness by experiments. To quantitatively assess the accuracy of pupil localization, we also propose a new experimental design to generate ground-truth pupil centers semi-automatically for a large amount of test video frames. The average localization errors of pupil centers of the proposed approach using several test videos are approximately lower than 1 pixel. The results of the proposed approach for circle fitting are better than those of Hough Transform in experimental comparisons.
In the second application for the construction of an Android App for real-time lane line detection, we, again, apply structural sampling to reducing the computational costs of lane line detection for smart phones with limited computing power. The proposed approach includes three processing steps: (1) block-based feature extraction, (2) vanishing point-based sampling, and (3) similarity measurement between land line structure and block samples. Preliminary experiments show that the proposed approach achieves 80% and 85% detection accuracy of lane lines in highway and street videos, respectively.
|
|---|
