Human Pose Estimation using Exemplar-based Particle Filter

We propose human pose estimation system which utilizes pose exemplars in the prediction step of particle filter. The particle filter assumes that state transition is surly represented by Markov chain model but our assumption is that if observation data and and prior reference dataset can be additionally used in prediction step, the prediction range can be more compact and precise at target posterior distribution. Therefore the proposed method have built exemplar pose dataset which is used in prediction step. From the proposal distribution of similarities of exemplars to observation data, particles are drawn and evaluated by likelihood score using chamfer matching. As a result, we choose best particle for pose estimation and which body pose is going to be an additional exemplar again in the exemplar dataset for improving performance of next prediction. Recursively the pose of 2D body parts are tracked by using exemplar-based particle filter. The proposed method efficiently estimates the pose of body parts as resultant images represent. 오치민, 상국초, 이칠우, Human Pose Estimation using Exemplar-based Particle Filter , 2013년 제25회 영상처리 및 R72이해에 관한워크샵, Feb. 2013