Full Body Motion Capturing Fusing Optical and Inertial Measurements



Art der Arbeit:

Masterarbeit von Schmeling, Marius Nicolas

Nowadays, the use of Motion Capturing (MoCap) systems is widely established. In addition to the high-end capturing systems used in the film industry, an increasing number of applications in the lower price segment can be found. However, using a commercial system for full body tracking results in relatively high costs. In this thesis, a hybrid system using both optical sensors of commercial Virtual Reality (VR) hardware and Inertial Measurement Units (IMU) is investigated. IMUs allow the user to track orientations with lower cost compared to commercial hardware. The main objective of the thesis is the fusion of different sensor data. An initial approach is to use the accurate optically tracked positions as anchor points due to the relatively poor tracking performance of the IMUs. The setup is as follows: each of the subject’s outer limbs (hands, feet, head) is tracked with a commercial sensor. In total, five commercial tracking points (Head Mounted Display (HMD), two handheld controllers and two additional foot trackers) and up to ten IMU points will be tracked with this system. Movement in between these sensors is to be determined using IMU data. The IMU sensors are located at the arms, legs, and at the upper body and lower body. These positions represent the crucial movement points of the human body. There are different approaches for the body pose estimation. Applying the rotations measured by the sensors directly to the bones is the most straight forward solution. As a more sophisticated approach, Inverse Kinematics (IK) can not only improve the accuracy but also require less IMUs since some sensors’ data is not required. With a fixed shoulder and wrist position for example, only a limited amount of arm positions are possible due to the fixed length of the arm. Thus, only one sensor, in this case on the upper-arm, is required and the sensor on the lower-arm is not used. Similarly, the leg position can be derived from the hip’s and the foot’s position. These principles can be transferred to the leg estimation. In the scope of this thesis the performance of three different pose estimation approaches is evaluated and discussed: 1. Transfer sensor rotations directly to the bones, 2. Using a 1-joint IK approach and 3. Using a 2-joint IK approach. It can be observed that the first approach generates the most natural body poses with smooth movement, but lacks in accuracy and suffers from drift. The 1-joint IK approach generates less natural body poses but archives higher accuracy. Furthermore, the 2- joint IK approach generates the most accurate body poses with more natural movement compared to the 1-joint IK, but its quality is highly dependent on precise calibration.