A novel approach for activity recognation based on local binary patterns

The most important technology that perceives and follows the movement of the human body is sensor based motion recognition technology. In recent years, sensor-based human motion recognition has attracted attention in the computer field due to its widespread use in everyday life and is a rapidly growing field of research. Motion recognition is performed by evaluating signals from various sensors placed in the human body. The success of motion recognition systems depends on the attributes obtained from the signals. The extraction of appropriate and effective attributes from the signals affects the accuracy of classification. In this work, a new approach has been proposed to obtain effective attributes from sensor signals. This method, One Dimensional Local Binary Pattern Model (1D-LBP) is a process for obtaining a new binary value as a result of comparison of the right and left neighbors based on the center sample value according to a certain neighborhood value of a value on the signal, and transforming the binary value into a decimal value. This method is a statistical metric that enables effective attributes to be obtained from various non-parametric signals. To test the proposed method, the signals of twelve physical activities of ten volunteers were used. A number of statistical attributes have been derived after applying the 1D-LBP method to the signals obtained from acceleration, gyro and magnetometer sensors. By using these statistical attributes, different machine learning methods and classifications have been performed. In this study, on the success of motion recognition; the effects of sensor types, sensor axes and extracted statistical attributes were also investigated. The highest success rate was %92,5 in all experiments.