New approaches for motion recognition from the wearable sensor signals

In recent years, there have been many different studies in the field of recognition, monitoring and differentiation of human movements. Two main approaches are generally used for activity recognition (AR). These are a vision-based approach and sensor-based approach. The sensor-based approach enables continuous monitoring of physiological signals that reflect the state of human movements, enabling the sensation of body movements. In the AR system, it is necessary to identify the attributes that best distinguish the movements from each other. After the feature selection is made, classification methods are carried out by machine learning methods. As the classification performance is reduced if the attributes are not selected well, attribute extraction plays an important role in the pattern recognition process. In this study, novel approaches to extracting features from sensor signals for AR are proposed. Nine different new approaches have been proposed for feature extraction. These approaches are: One-Dimensional Local Binary Patterns (1D-LBP), Average Based One-Dimensional Local Binary Patterns (AB-1D-LBP), Median Based One-Dimensional Local Binary Patterns (MB-1D-LBP), Multi-Dimensional One-Dimensional Local Binary Patterns (MD-1D-LBP), Weighted One Dimensional Robust Local Binary Patterns (W-1D-LBP), Neighborhood Based One-Dimensional Local Binary Patterns (NB-1D-LBP), One-Dimensional Local Binary Patterns with Shift (S-1D-LBP), Ternary Patterned One-Dimensional Local Binary Patterns (TP-1D-LBP) and Reduced One-Dimensional Local Binary Patterns (R-1D-LBP). The proposed approaches are developed based on the One-Dimensional Local Binary Patterns method developed based on the Local Binary Patterns method used frequently in image processing applications. The proposed AR system consists of six stages. In the first stage, sensor signals of various physical movements of the people were provided. In the second stage, more significant signals have been obtained by applying the feature extraction methods developed in these sensor signals. In the third stage, histograms of the significant signs obtained were generated. In the fourth stage, attributes were obtained from generated histograms. In the fifth stage, the obtained attributes are classified by Random Forest Algorithm machine learning method. In the sixth stage, the comparison of feature extraction methods developed according to various criteria was made. According to the Random Forest algorithm, the highest success rate was 95,83% and the R-1D-LBP method was used. In this study, it has been observed that AR treatment is performed with high success rates by using all new approaches and sensor signals.