NHK STRL at TRECVID 2009: Surveillance Event Detection and High-Level Feature Extraction.
Masaki Takahashi, Yoshihiko Kawai, Mahito Fujii, Masahiro Shibata, Noboru Babaguchi, Shin’ichi Satoh · TRECVID · 2009
NHK Science and Technology Research Laboratories participated in two tasks at TRECVID 2009: surveillance event detection task and high level feature extraction task. For surveillance event detection tasks, we targeted four events: PersonRuns, PeopleMeet, O b jectPut, a n d OpposingFlow. The proposed method detects human regions using HOG descriptor and SVM classifier and tracks human regions using 2D color histograms. It recognizes each event based on trajectories of people and optical flow. For the high-level feature extraction task, we calculate feature vectors based on local features and global features, and then classify key-frames using a ball vector machine (BVM). The proposed method uses both SIFT feature and SURF feature to get various kinds of local feature points and descriptors. Features such as color moments, wavelet texture, local binary patterns, and face appearance are used as global features. We adopt the BVM method to reduce the computational cost of training.