Scale Invariant Feature Transform의 연산과 응용을 실제시간에 수행

다국어 초록 (Multilingual Abstract)

Noways,image matching is used to solve many problem in computer vision, including object or scene recogniton,3D structure from multiple images, stereo correspondence, and motion tracking. In recent years, an approach has been proposed to generate a set of image features.This approach has been named the Scale Invariant Feature Transform;it transforms image data into scale-invariant coordinates relative to local features. The SIFT algorithm combines a scale invariant region detector and a descriptor based on the gradient distribution in the detected regions. The descriptor is presented by a 3D histogram of gradient locations and orientations. Those descriptors(local features)are very distinctive and invariant for image scaling or rotation.SIFT keypoint is used in many object recognition applications because of this propery.
However, SIFT algorithm is a complex algorithm. To apply SIFT in multimedia applications, it is necessary to find a scheme to implement the algorithm in real-time.
In this thesis, I used a parallel approach to implement the SIFT algorithm. In my implementation,a single instruction stream multiple data stream pixel processor is used. By using the SIMD pixel processor system,the available parallelism of the SIFT algorithm can be exposed fully.Major results showed that we can apply this new approach to implement the SIFT algorithm in real-time.

목차 (Table of Contents)

• CHAPTER 1 INTRODUCTION = 1
• 1.1 Problem Overview = 1
• 1.2 Thesis Objective = 3
• 1.3 Thesis Outline = 4
• CHAPTER 2 SIFT ALGORITHM = 5
• 2.1. SIFT Algorithm Description = 5
• 2.1.1. Detection of scale-space extrema = 6
• 2.1.2. Local extrema detection = 9
• 2.1.3. Accurate keypoint localization = 10
• 2.1.4. Eliminating edge responses = 12
• 2.1.5. Orientation assignment = 14
• 2.1.6. The local image descriptor = 16
• 2.1.7. Descriptor representation = 16
• 2.2. SIFT Evaluation = 19
• 2.3. Applications in multimedia = 20
• 2.3.1. Object recognition using SIFT features = 20
• 2.3.2. Robot localization and mapping = 21
• 2.3.3. Panorama stitching = 21
• 2.3.4. 3D scene modeling, recognition and tracking = 22
• 2.3.5. 3D SIFT descriptors for human action recognition = 22
• CHAPTER 3 PREVIOUS WORKS = 24
• 3.1 PCA-SIFT: A More Distinctive Representation for Local Image Descriptors = 24
• 3.2 A High Real-Time and Robust Object Recognition and Localization Algorithm = 26
• 3.3 Fast approximated SIFT = 29
• CHAPTER 4 IMPLEMENT THE SIFT ALGORITHM IN SIMD PROCESSOR = 33
• 4.1. SIMD Pixel Processor Architecture = 33
• 4.2. Implement SIFT algorithm on SIMD Processor = 35
• CHAPTER 5 PERFORMANCE ANALYSIS = 39
• CHAPTER 6 CONCLUSIONS = 44
• REFERENCE = 45