Geometric Revise using Distance Transform for Pose Estimation

Vision-based pose estimation is highly active research area due to the various applications. We propose the method using both inverse kinematics and distance transform for more realistic pose estimation. First, we roughly estimate the three dimensional pose from the partial area of human body. This process is composed of human detection, skin color detection, labeling, morphology, blob tracking and linear triangulation method for 3D reconstruction. Secondly, we estimate the similar pose of the target human from the intermediate joints generated by inverse kinematics. Before applying the numerical calculation of inverse kinematics, we use an analytical method to reduce the number of iteration, and which brings less computational cost than that of the algorithm which utilizes only numerical method. Finally, we optimize the temporal state of joints which is estimated by inverse kinematics with the distance map of body silhouette. Through this optimization process, we can obtain more precise pose solution than those poses calculated by only inverse kinematics theory. The experimental results show that our proposed method surely complements the defect of inverse kinematics algorithm with the analysis of the constrained DOF. Through above process, the proposed system is confirmed to do the more precise pose estimation with the geometric revise. Y. Seo, C. Oh, D. Song, K. Bae, C. Lee, Geometric Revise using Distance Transform for Pose Estimation, 16th Korea-Japan Joint Workshop on Frontiers of Computer Vision (FCV 2010), pp. 277-282, Feb. 2010