PEOPLE DETECTION USING FEATURE VECTOR MATCHING

다국어 초록 (Multilingual Abstract)

This paper presents a novel method to detect people in real time from a video sequence taken from a single fixed camera. Foreground images are obtained from a sequence of operations: background differencing, thresholding, and morphological operation. We make use of quantized edge orientation to represent shape of human because human has particular edge orientation distribution on each of body parts. Then, a feature vector derived from the edge orientation map is computed in each of body parts of foreground images. The dataset, composed of sets of feature vectors, is used for matching with feature vectors computed from the current image of the input sequence. We use the Knearest neighbors to match the feature vectors. The proposed people detection method uses a proper feature vector that represents human edge orientation distribution, and employs simple matching steps for real-time processing. Experimental results with a number of test sequences demonstrate the effectiveness of the proposed method.

목차 (Table of Contents)

• ABSTRACT
• 1. INTRODUCTION
• 2. FOREGROUND REGION DETECTION
• 3. PEOPLE DETECTION
• 4. EXPERIMENTAL RESULTS AND DISCUSIONS
• 5. CONCLUSION
• Acknowledgment
• 6. REFERENCES