IMPROVEMENT OF ACCURACY AND SMOOTHNESS LEVEL ON SKELETAL-JOINT KINECT DATA BY USING KALMAN FILTER ALGORITHM

• Main Author: Tities Ginalih, Cecilia
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/54406
• Institution: Institut Teknologi Bandung

Kinect sensors can be used in robotic field to move a robot arm so that the interaction between humans and robots becomes more natural. Kinect sensors the human body by tracking the joints of the body which is called skeletal tracking. Skeletal tracking data consists of 20 data joints which positions are stored in 3D coordinates. Although skeletal tracking data has a good accuracy, there is often noise in the form of unwanted data jitter or data spikes. Noise contained in the skeletal tracking data can be overcome by a filtering process. The filter algorithm that is widely used in smoothing the Kinect data is the Kalman Filter. Apart from applying the KF algorithm, this study will also compare the KF and EKF estimation results. The research begins by designing the filter characteristics which are state vector variables and the R and Q parameters for the KF and EKF algorithms respectively. After finding the R and Q values, both filters are applied to the joint for two types of poses, static and dynamic poses. In static poses, the level of accuracy and smoothness are obtained from the mean values of MAE and SE (KF 0.012461 and 0.000148, EKF 0.009157 and 0.000063, Kinect 0.01238 and 0.00018). From the calculation results, it was found that the performance of EKF was better with lower MAE and SE values than KF and Kinect. Meanwhile, the implementation of EKF in dynamic poses is analyzed graphically because there is no database that can be used as a reference value. EKF results in dynamic poses have the potential to reduce the requirement for position data to identify gestures. The R and Q combination trial was conducted to see the effect of the difference between R and Q in static and dynamic poses. In this trial, the best parameter values for Q are 10- 6 and R is 10-4 . The results of the filter estimation of the two poses indicate a reduction in noise on the joint position while maintaining the gesture data.