Improving human kinematic data using kalman filter

The detection, recognition and storage of human movement; It is of great importance for measuring or improving the accuracy of the work done. Some analyzes are made to examine the behavior of people during their movements. The first and most important stage is the digitization part of the movement. While obtaining numerical data, the proximity of the data to the actual measurements is extremely important. With the development of sensor technologies, in addition to the sensors worn on the body and depth map sensors are allowing 3D modeling of human movements, also supported with image processing software. Measurement errors can be observed in motion capture sensors as in many electronic systems. Different filters and algorithms must be developed in order to eliminate these errors and reduce them to an acceptable level. Sometimes the data flow from the sensors may be interrupted for a short time. In this case, estimation algorithms should be used. Kalman Filter (KF) is a powerful algorithm that can predict next data based on previous data. In this study, the real body lengths of the individuals were measured, the same lengths were calculated according to joint coordinates obtained from the human motion sensor and compared with their actual body lengths. Average absolute error rates were calculated for these two values and it was observed that the error rate was high. In order to minimize the measurement errors obtained and to reach more realistic measurements, a special filter should be designed for this system. One of the most commonly used filters to optimize measurement errors is Kalman Filter. Although KF is very powerful in correcting errors in linear systems, it can also produce solutions for nonlinear systems with extended Kalman filter (GKF). The aim of this study is to design a Kalman and Extended Kalman filter for real-time measured data from a human motion sensor, and produce values much closer to real data. Since human movements have a non-linear structure, it has been observed that measurement errors decrease to acceptable levels when GKF is applied. Mean absolute percentage error was used for performance evaluation; It was calculated as 19.962% error rate in measurements made with Kinect device, 14.002% error rate using Kalman filter, and 7.693% in Extended Kalman filter. In this way, a human motion capture system, which is easy to adapt, saves time and money, and can perform real-time and accurate measurements, has been developed through only one sensor and the software developed without taking actual measurements of the persons.