- Author: Li Youhong, Software Engineering, College of Software
- Adviser: Dr.Guo Shihui， Associate Professor
- Contact: liyouhong@stu.xmu.edu.cn
mocap dataset; deep learning; temporal relation; complex tasks; geometrical lengths; posture estimation;
Inverse Kinematics (IK) is a long standing problem in the fields of robotics control and character animation. The main challenge lies in the redundancy, where an infinite number of body configurations may reach the same position of end-effector. Selecting the appropriate one from the large repertoire of candidates is an open question. It is particularly challenging to identify the natural posture for humanoid characters since we are most familiar with human motion and highly sensitive to subtle details. This paper addresses the problem of Inverse Kinematics with deep neural network, trained with so-far the largest human mocap database. We identify the critical temporal correlation between input and output motion frames and compare systematically the performance of nerual network training with different parameters. Given the challenge of multi-solution in the IK problem, the trained model selects the pose which is most consistent with the pose by the real performer. This consistency is validated by the comparison with the benchmark database. A denoising filter is proposed to further improve the prediction results based on the actual performance of the network. The trained model is adaptable to complex tasks, such as basketball dribbling and ballet dancing, and to characters of different geometrical lengths. We do not assume the knowledge of the accurate limb lengths and skip the requirement of manual setup of joint limits. This eliminates differences between individuals and allows our method to be directly used in the problems of posture estimation.