This paper introduces a new real-time human motion parameters capturing method using Kinect. It consists of five modules. First, the hybrid action type classifier categories human body motion into four different action types. Second, for each action type, there is a body part (BP) classifier which segments the human silhouette into 16 BP regions of which the centroids become the BP joints. These BP joints are linked to represent the human body skeleton. Third, an action type validation process verifies the identified action type. Fourth, the partial occlusion recovery method relocates the occluded BP joints. Fifth, the offset compensation process fine tunes the positions of BP joints and then validates the compensation results. The major contributions of this paper are hybrid action type classification and correction, offset compensation, and partial occlusion recovery. The experimental results show that this method can estimate human upper limb motion parameters in real time accurately and effectively.

This paper proposes a real-time upper human motion capturing method to estimate the positions of upper limb joints by using Kinect. For human articulated motion capturing, the body part self-occlusion is a nontrivial problem. The system consists of hybrid action type recognition, body part segmentation, and offset compensation. The hybrid action type classifier consists of Adaboost and Random Forest classifier. The major contributions of this paper are offset compensation and self-occluded joint recovery. The offset is the difference between the output and the ground truth. The offset compensation is proposed by correcting the estimated locations of the joints. For different user action type, we train an appropriate offset classifier for offset compensation. Finally, we propose a postprocessing to justify the effectiveness of the offset compensation.