Real-time video signal processing methods for feature extraction and analyses

In this thesis, two topics of computer vision are presented. The first topic presents a real-time hand segmentation algorithm that is based on background subtraction and color information. Unlike utilizing one method to segment the hand, the strong points of two methods help each other to segment the hand clearly and robustly. A hand is extracted from a video sequence by background subtraction where unit gradient vectors (UGVs) are used instead of image intensities. Since the UGVs are invariant to change in illumination, the UGV-based background subtraction is more stable dynamic lighting conditions. The color information improves the UGV-based background subtraction result. Therefore, the integration of two results can segment a hand under various lighting conditions robustly. Finally, this algorithm is implemented to a low-cost embedded board Raspberry Pi for a real-time application system. The second topic presents a novel technique for classifying several camera operations in videos. The camera operations have important roles in different research areas such as video indexing, video stabilizing, video encoding, video processing, and video content analysis. To perform the classification, 2D motion vector (MV) fields are firstly measured by block-based motion estimation. Then, the 2D MV fields are described as 2D MV histogram in polar coordinates. The histogram shows how many MV magnitudes and MV orientations in each video frame. These two MV features are utilized simultaneously to classify a total of 16 camera operation classes including single camera operations and combinations of two camera operations. The proposed method can achieve a processing time of 5 – 10 millisecond per frame for a low-resolution video, while it takes 40 – 80 millisecond per frame for a high-resolution video.