능동시각을 위한 모션정보 기반 선택적 주의집중 시스템 = Selective Visual Attention System Based on Motion Information for Active Vision System

다국어 초록 (Multilingual Abstract)

How to detect meaningful video representation becomes an interesting topic in various research communities. Visual Attention System is proposed to detect “Region of Interesting” from input video sequence. Generally the attended regions correspond to visually prominent object in the image in video sequence.
In this paper, we suggest a novel method to improve previous approaches based on spatiotemporal attention modules. We propose a new integration model to make use of depth information in addition to spatiotemporal features. As a result the proposed method compensates typical approaches for inaccuracy improving performance.
Motion is important cue when we derive temporal saliency. But noise obtained during the input and computation process deteriorates accuracy of temporal saliency. To remove the noise from motion information we exploited the result of psychological studies. The function of "double opponent receptive field" and "noise filtration" in Middle Temporal area is simulated. We also applied "FlagMap" on each frame to prevent "Flickering" of global-area noise. These considerations make the proposed system detect the salient regions with higher accuracy while removing noise effectively.
Depth information is used to detect salient regions. Typical systems get problems in determining the saliency if several salient regions are partially occluded and/or have almost equal saliency. However, the proposed system is able to separate the regions with high accuracy. Spatiotemporally separated prominent regions in the first stage are prioritized using depth value one by one in the second stage.
The proposed method has been applied to several image sequences. Experiment result shows that the proposed method can describe the salient regions with higher accuracy than the previous approaches do.

목차 (Table of Contents)

• 1. 서론………………………………………………………………………1
• 2. 시각주의의 개념과 주의영역 추출방법 및 알고리즘 …………2
• 2.1 공간적 주의 ……………………………………………………3
• 2.2 시간적 주의 ……………………………………………………6
• 2.3 공간적 현저도와 시간적 현저도의 결합 ……………………8
• 2.4 현저도를 구하기 위한 기반 알고리즘 ………………………11
• 2.4.1 색상…………………………………………………………11
• 2.4.2 밝기…………………………………………………………13
• 2.4.3 블록매칭 알고리즘 …………………………………………15
• 2.4.4 방향 …………………………………………………………17
• 2.4.5 활동량의 비교와 맵 결합 …………………………………24
• 2.4.6 적응적 결합 …………………………………………………26
• 2.5 한계점 …………………………………………………………27
• 3. 제안 ………………………………………………………………29
• 3.1 우선순위 부여와 영역 분리 …………………………………29
• 3.1.1 깊이맵………………………………………………………31
• 3.1.2 깊이맵의 활용 ………………………………………………32
• 3.2 효과적인 노이즈 제거 …………………………………………33
• 3.2.1 이중대립수용야의 반응과 노이즈 여과 …………………33
• 3.2.2 깜빡임의 제거 ………………………………………………36
• 4. 실험결과 …………………………………………………………38
• 4.1 우선순위 부여와 영역 분리 …………………………………38
• 4.2 노이즈 제거 ……………………………………………………43
• 4.3 결과 비교………………………………………………………47
• 5. 결론 ………………………………………………………………54
• 참고문헌 ……………………………………………………………55