[11.23] Recognizing gestures for virtual and real world interaction

ICVS 2009
MIT CSAIL, Toyota Research Institute

내용
Stereo → Articulated ICP → Recognition.
model을 ICP로 3D data에 fitting할 때 multi-hypothesis framework 사용. N = 4 hypothesis is enough.
recognition은 bag of features (BOF) 기반이며, temporal descriptor로 codewords의 histogram을 사용. histogram을 SVM으로 training. 물론 classes는 different gestures.
비교는 HMM과 Conditional Random Fields (CRF).
실시간으로 동작한다고.